Update now updates, however you installed wisp.Update to X installs the new version and restarts wisp for every install route: npx, npm install -g, the install.sh script, and now Homebrew. On a Homebrew install it used to run npm install -g and then restart the old keg. Now it runs brew upgrade m0xoo/tap/wisp-agents and restarts from the new keg. If the install leaves wisp on the same version (the tap doesn't have the release yet, or npm installed into a different folder), the button says so instead of restarting into the old version and then reporting that wisp didn't come back. A git clone of the repo no longer shows the button or tells you to run git pull. GET /update no longer returns auto, and the method can now be brew.
How to get a missing harness, where you look for it. Settings → Agents now shows, for each harness wisp didn't find, what to do on that computer's OS: a one-line tip (e.g. that the ChatGPT app is enough for Codex on macOS and Linux but not on Windows, or that the Antigravity and Cursor editors don't include their CLIs), the install command to copy with one click, and a Docs link. In the harness picker, hovering a greyed-out harness shows the same tip, and clicking it (or arrowing onto it) shows the steps in place of its models, with Re-check (and Install for opencode). It still can't be picked. API: unavailable harnesses in GET /agents carry setup: { tip, command?, url }; adapters provide it through an optional setup(platform).
Codex and Claude Code from their desktop apps, no CLI install needed. If codex isn't on PATH, wisp uses the Codex CLI the ChatGPT desktop app ships: ChatGPT.app/Contents/Resources/codex-cli/bin/codex on macOS (or Codex.app/Contents/Resources/codex from the older Codex app), in /Applications or ~/Applications, and /usr/lib/chatgpt/resources/codex from the Linux .deb. It shares ~/.codex, so the app's sign-in carries over. The Windows app is a Store package whose files can't be run from outside it, so on Windows Codex still needs the CLI. If claude isn't on PATH, wisp uses the newest copy the Claude desktop app downloads for its Code tab, on macOS and Windows (<app data>/Claude/claude-code/<version>/…, including the Store installs' folders). That copy is signed in only while the app runs it, so run it once and /login. Both adapters also look in the CLIs' own install folders (~/.codex/packages/standalone/current/bin, %LOCALAPPDATA%\Programs\OpenAI\Codex\bin, ~/.local/bin, ~/.claude/local), which an app started from the Dock or Start menu may not have on PATH. A WISP_CODEX_BIN or WISP_CLAUDE_BIN override still wins.
Changed
The website shows moving a task to another computer or agent. A new landing-page section, Start on your laptop. Finish on the dev box., has a working copy of a task card's follow-up box: pick a computer under Runs on and an agent, press ↵, and it packs the work, hands the history over and carries on there, the way the app does. It plays once by itself when it comes into view; click anything to try it. A new guide, Move a task to another machine or agent, mid-way, covers doing it by hand with git bundle, what wisp moves, what each agent is told when you switch harness, model or computer, and the CLI and API. The landing FAQ answers it too, and Machines is in the landing nav. In Also in the box, the Machines tile became Several folders, since machines now have their own section. The landing nav no longer wraps onto two lines between 960 and 1200px wide; it drops Control and CLI there.
Fixed
The site builds the same on Windows. With a CRLF checkout, the integrations page kept its markdown title as a second <h1> and the changelog page repeated each release heading, because scripts/site.mjs matched \n line ends only. It now normalizes line ends when reading docs/integrations.md and CHANGELOG.md, and the site test passes on Windows CI.
Done and Archived count every task and load more as you scroll. Their headings always said 60 once you had more than that, and the older tasks never showed. The headings now count them all, the lists show 60 cards at first and add the next 60 as you scroll near the end (or click Show 60 more · N left), and opening an older task from a notification or ⌘K shows the cards down to it.
Auto's Balanced goal runs frontier models. When you have a frontier model installed and allowed (Opus, Codex's default, Gemini Argon, Grok 4.x), Balanced now picks it, at the harness's default or medium thinking, instead of Sonnet or Haiku. Best stays on frontier models at high or x-high, so the thinking level is what sets the two apart. Balanced steps down to an everyday model (Sonnet, never Haiku) only while every frontier model is at its limit, so the task still runs, or when you have none; small models' wins no longer pull Balanced to them after 20 outcomes. Cheap is unchanged. For models with the level in their name (Antigravity's gemini-4-argon-low / -medium / -high), the goal now picks the variant: Cheap low, Balanced medium, Best high, and the other variants count as the same model.
Full transcript, when you want it. Turn on Settings → General → Full transcript and every card shows its run as a timeline: the agent's messages in full, every step it took (commands, reads, edits, searches, questions), and at the end its presentation, rendered as blocks as usual. Click a step to see the whole command and what it printed, the diff that one edit made, or the full error. Long runs of steps fold their middle, and Jump to the result skips to the end. While a task runs, the timeline grows live above a one-line status bar with Stop. Any card can switch either way with its Transcript button or V (also in ⌘K), and keeps that choice. Off by default. Agent messages are now kept whole (they used to be cut at 600 characters), and so is the output of each command, search, web or MCP call, up to 2,000 characters (its start and end), from Claude Code, Codex, opencode, Cursor and Grok. Turns from before this change show their steps without output.
Every project runs on every machine. Projects are no longer tied to one computer. Pick any project, pick a machine under Run on, and launch: if the machine has no checkout of it yet, wisp finds one there (a project or a git repo in the usual folders with the same origin) or clones the repo's origin into the machine's clone folder, then remembers it. The launcher shows one row of projects for all the places you picked, the card says "Cloning … on devbox" while it does that, and the Projects dialog lists every project with the places it's checked out on. A folder you add on a machine that's the same repo as a project you already have joins that project instead of becoming a second one. Repos without a git remote can't be cloned this way; wisp says so.
Move a task to another computer with a follow-up. The composer has a new Runs on button (shown once you have a machine). Pick another computer and your next follow-up moves the task there: its branch, uncommitted and untracked changes travel as a git bundle and land in a worktree on the same branch, and the earlier turns come along as history, so the harness there is told what happened before it continues. The old session is archived with a "Continues on …" link; the new card shows "From …" and tags each earlier turn with where it ran. A running task has to stop or finish first. API: "machine" on POST /sessions/:id/reply ("" for this computer).
Rerun on another computer. The rerun dialog has a Run on picker; the rerun uses that computer's checkout of the project, cloned there first if needed.
/btw side questions. Type /btw did you update the docs too? in a card's follow-up box to ask while the agent works, without interrupting it. A separate read-only run answers it, and the answer shows on the card under the turn you asked during. Claude Code, Codex and opencode answer on a copy (fork) of the agent's conversation, so they know what the run did and why; the task's own conversation never sees the question. Cursor, Antigravity and Grok, and CLIs too old to fork, get wisp's record of the task instead: requests, reports, actions so far and files changed. It also works when the task isn't running, and it adds no turn. Dismiss an answer with its ×; dismissing one that is still running stops it. Typing / in the box opens a small command menu (/btw, /feedback), and a working task shows a quiet Tip: /btw asks without interrupting next to Show all activity. Click it to start one in the box. It goes away after your first /btw, or when you hide it with its ×. CLI: wisp btw <id> "question" prints the answer, and wisp reply <id> "/btw …" works too, as does a /btw reply from Telegram. API: POST /sessions/:id/btw, DELETE /sessions/:id/btw?item=<id>. SDK: wisp.btw(id, text).
Changed
No folder is picked by default when wisp has no folders. With no projects added, the launcher now starts on No folder, so a prompt runs right away in an empty scratch directory instead of opening the folder picker. Adding a folder picks it, as before.
The website is now trywisp.dev, with more pages for search engines to find. Canonical, Open Graph, structured-data, robots.txt, llms.txt, the install script and the npm and Homebrew homepage use the new domain. The site gained a page per harness (Claude Code, Codex, Cursor CLI, opencode, Antigravity CLI, Grok Build) that shows how wisp runs that CLI: the flags behind each permission mode, how it asks you questions, /btw and the take-over command. An agents overview compares them, and the integrations guide and changelog are now pages on the site instead of GitHub links. scripts/site.mjs (npm run site) builds these pages, a 404.html and a sitemap.xml with each page's last-commit date. The Pages workflow runs it before each deploy, redeploys when CHANGELOG.md or an adapter changes, and pings IndexNow (Bing and friends) afterwards. The harness logos on the landing page link to their pages.
Fixed
The Auto-fix PR conflicts switch works on its own. In Settings → General it only responded after you had clicked Worktree per task in the same visit.
Chevrons inside Done cards point the right way. Showing the Done list rotated every chevron in its cards too, so collapsed cards looked open.
The rerun dialog's Rerun button no longer slides off the edge. With the Run on picker added, the bar under the prompt (place, mode, harness) was wider than the dialog and pushed Rerun out of view. The dialog is a little wider, and when space runs short the harness name folds first (its logo stays), then the model; on narrow windows the place and mode buttons show only their icons.
One task across several folders. A task can now work in more than one project at a time, for example an API and the web app that uses it. In the launcher, ⌘-click (Ctrl- or Shift-click) a second project chip, or click the + on its corner or next to it in the project list. One agent starts in the first folder and can read and edit the others: Claude Code and Codex get them as --add-dir, and every harness is told the paths. Each git repo gets its own worktree. The card reads api + web and shows up under either project's filter. Ship it opens a pull request in each repo that has changes, and Merge all merges them together and syncs each local base branch. CLI: wisp "…" --also ~/code/web (repeatable). API: "also": [folders] on POST /sessions. MCP: also on wisp_launch.
Fixed
The launcher's project row no longer throws while it isn't laid out. Refitting the chips of a launcher that had just been painted over (or folded into the dock) read the position of a chip that wasn't there.
Fix conflicts button on every conflicting PR. When a PR has merge conflicts, its row on the card now has a Fix conflicts button (with a merge icon) that sends the task a follow-up to merge the base in and resolve them; wisp pushes the result to the PR. The row switches to "Fixing merge conflicts" as soon as you click, and the follow-up starts without waiting on a fetch first.
Changed
PR conflicts are no longer fixed without asking. wisp used to send the merge follow-up on its own as soon as a PR conflicted. Now it waits for your click on Fix conflicts. To get the old behaviour back, turn on Settings → Auto-fix PR conflicts.
PR status updates faster. A PR whose checks are running, whose mergeability GitHub is still working out, or that was just pushed is now checked every 10 seconds instead of every minute, so a passing check or a cleared conflict shows up quickly. The refresh button on the PR row spins while it checks.
Auto no longer runs Best tasks twice by default. Best of two (two picks side by side, the better one marked) is now off unless you turn it on in Settings → Auto → Best of two, or with PATCH /api/router{ "best2": true }. With the Best goal, a task now runs once, on a single pick. Your manual vs Compare in the launcher is unchanged.
Auto's Best goal runs frontier models only. Best now always picks a frontier model (Claude Opus, Codex's default, Grok 4, Gemini Argon), never an everyday or small one like Sonnet or Haiku, however well those did on your past tasks. Before, that was only true for your first 20 outcomes. If every frontier model is busy or at its limit, Best waits instead of stepping down: best-of-2 runs a single pick, and a move at a usage limit stops with a note. Only when you have no frontier model installed (or allowed) does Best use the strongest one you have.
An open card stays open when its task finishes. It used to fold shut if you had scrolled away or switched windows while it worked; now it lands in "Needs you" still open, with the result ready to read.
Fixed
Right-click menus no longer get cut off at the bottom of the window. The task, project and queued-message menus measured themselves mid pop-in animation, while still scaled down, so near the bottom edge the last item ("Delete session") ended up below the fold. They now measure their real size and stay fully on-screen.
Auto picks the thinking level too. Auto's options are now harness · model · thinking level (claude:opus:high), and the goal sets the range of levels: Cheap runs low (or minimal), Balanced the harness's default or medium, Best high or x-high. Within that range Auto learns which level works for each kind of task, the same way it learns models. A model's record at one level counts half at its other levels. Harder thinking starts out a little more likely to work and costs more, and then your outcomes decide. Compare runs and hand-overs use a different model, not just a different level. A rate limit rests the model at every level. A follow-up that changes the level counts as a miss for Auto's pick. Settings → Auto lists the levels it has a record for. Harnesses and models without levels run at their default, as before.
Changed
Shell commands read as English in the live ticker. Instead of Running sed -n 15,50p server/a.js, a task now shows "Reading lines 15–50 of server/a.js". wisp knows the common file, search, git, gh, npm/pnpm/yarn, Python, Node, cargo, go, make and docker commands, follows && chains and pipes (git status && git diff --stat → "Checking git status, then summarising changes"), and looks inside bash -lc '…'. Works for every harness; a command it doesn't know still shows as Running …, and the exact command stays in the activity detail.
The published 0.18.0 package was missing the Auto checks. It was built before they merged, so npm i -g wisp-agents got none of the entry below. 0.18.1 ships them.
Added
Auto checks the work before it reaches you. When an Auto run says it's done, wisp runs the project's tests itself: npm test, pytest, go test, cargo test, make test, and so on. You can set the command, or turn the check off, per project in Settings → Auto → Open. A failure only counts if the test passed before: wisp compares with the base commit and runs a failing suite twice.
Repair: tests that used to pass now fail, so they go back to the same agent. One round by default; the turn shows as Sent by wisp.
Hand over: still failing, so the task runs again on a stronger tier (small → everyday → frontier) in its own worktree. Both runs show as one task.
Usage limits don't count against a model. When a harness says it's at its limit (Claude's session limit, a rate limit, no credit left), the run doesn't count as a failure. That harness, or just that model for a rate limit, rests until its limit resets, and the task moves straight to the best pick that can run now (Moved on). Settings → Auto → Open shows what's resting and until when.
Quality first until Auto knows you. Until Auto has 20 of your own outcomes, Balanced only picks everyday and frontier models, and Best only frontier ones. The shipped starting estimates count for less (at most 4 tasks each).
Best of two: with the Best goal, two picks work side by side and wisp marks the better one. Tests decide first, then a cheap model reads both diffs.
Scout: before a bigger task starts, a cheap model looks around the repo and tells the agent which files matter and how wisp will test.
Each turn gets wisp's own tests block (Run by wisp). If the agent said "done" and wisp's run disagrees, the turn says so and the card shows Partial.
Cards show Tests / N failing, Pick and Handed on.
Your own follow-up or Stop cancels whatever Auto was about to do next. Settings → Auto turns each step on or off. API: PATCH /api/router with check, repairs, escalate, best2, scout; PATCH /api/projects/:id with testCommand.
Merge conflicts on shipped PRs fix themselves. When something lands on main and your shipped PR no longer merges, the task used to sit there until you noticed on GitHub. Now wisp sees it on its PR check (every minute, or when you press refresh): GitHub reports the PR as conflicting, and wisp confirms it locally with git merge-tree against a fresh fetch of the base, so a stale GitHub answer can't start a run. The task then gets a follow-up asking it to merge origin/main in (no rebase, since the branch is pushed), resolve each conflict keeping both sides' intent, run the cheap checks and commit. When that run ends, wisp checks the merge is complete: no unmerged files, the base is in the branch, no leftover <<<<<<< markers. Then it pushes to the PR, and the card shows the usual Push → PR track. While it works, the PR row reads Fixing merge conflicts and hides Merge. If the agent stops halfway, nothing is pushed: the card comes back unread with Conflicts not fixed and the reason, plus a Fix conflicts button to try again. wisp tries once per PR head and at most 3 times per task, so a base that keeps moving can't run up the bill. On by default; turn it off in Settings → General → Fix PR conflicts (defaults.fixConflicts: "off"). API: conflict on the PR, conflictFix on the session summary, POST /api/sessions/:id/fix-conflicts.
Tasks without a folder. The launcher has a No folder chip next to your projects, for general asks that have nothing to do with a repo (explain a concept, draft a message, research something). The agent runs in an empty scratch directory under ~/.wisp/scratch, with no worktree, and the task shows under General. The folder is not added to your projects. The Worktree toggle hides while No folder is picked, and deleting the task deletes its scratch folder. API: POST /api/sessions {"prompt","scratch":true}.
Add a project from a GitHub URL. Paste https://github.com/owner/repo (or git@github.com:owner/repo.git, a link to any page inside the repo, or any other git URL) into Projects and the button turns into Clone. Click it and the dialog closes: the launcher already points at the repo (a chip with a spinner, cloning…) and the prompt has focus, so you write the task while it clones. Send it before the clone finishes and the task card shows Cloning owner/repo, then starts in the clone. If the clone fails, a toast says why, the launcher goes back to the project you had picked and your task comes back into the box. A feed filtered to another project switches to All, so the launcher isn't pulled back to that project. GitHub repos go through gh repo clone when gh is installed, so private repos work with your gh login; otherwise plain git clone. Repos land in ~/code (or the first of ~/dev, ~/Developer, ~/projects, ~/src, ~/repos… that exists). Click the path in the hint to change it (defaults.cloneDir). If the folder there already is that repo, wisp reuses it; another folder with that name gets a -2 suffix. When gh is signed in, the dialog also lists Your GitHub repos, newest push first and minus ones already added, so cloning one is a single click. On a machine tab the clone happens on that machine. CLI: wisp add https://github.com/owner/repo. API: POST /api/projects {"repo":"…"}; a URL in path works too.
Changed
Project chips stay on one row with tens of projects. The launcher's in row used to show every project as a chip (30 projects meant 6 rows on desktop and 14 on a phone), and the top bar cut off whatever didn't fit with no way to scroll to it. Now both show your most recent projects as chips, and the rest go behind a +N chip. Fewer chips show when the space is narrow, and a chip is never cut in half. The project you picked, the one you're filtering by, and projects with tasks that need you always keep their chip. +N opens a searchable list of every project: type to filter by name or path, use ↑/↓ and Enter to pick. It also has No folder and Other folder… in the launcher, and All projects and Manage projects… in the top bar. In the top bar, +N shows a badge counting the tasks that need you in the projects it hides. When a phone is too narrow even for All, the top bar shows one chip with the current filter that opens the list.
Ship it no longer holds you in its dialog. Clicking Commit, push & open PR used to leave the dialog open with a disabled Shipping… button until the push and gh pr create were done. Now the dialog closes right away, and the card's PR row (just above the composer) shows a live track: Commit → Push → Open PR, with a spinner on the step that's running and a check on each one that's done. The server reports each step as shipping on the session summary, so other tabs see it too, and the card head shows a Shipping pill while the card is collapsed. When the PR is up, the row turns into the PR row with a short highlight, and a toast links to it. If a step fails, the row turns red and says which step failed and why, with Retry, which reopens the dialog with your commit message and description still filled in. A second ship while one is running is refused with 409.
Fixed
The launcher is there from the first frame. On start, the launcher used to sit as an empty strip for one to five seconds, because the page waited for harness detection (GET /api/agents runs every CLI, and opencode lists its models) before drawing it. Now index.html draws the launcher's shape with a prompt box that already has focus, so you can start typing before the scripts load. The full launcher fills in about 50 ms later, at the same height, and keeps what you typed. Harnesses come from the last list seen (kept in the browser) and refresh when detection answers. GET /api/agents now answers from the last detection right away and checks again in the background, sending an agents event if anything changed. Warm calls dropped from about 1.2 s to 2 ms. ?fresh=1 and Re-check still wait for a full detection. Empty feed sections no longer flash their headings before the feed loads.
Approve what an agent's permissions block, from wisp. In Edits mode a coding agent used to be refused quietly when it wanted something its permissions don't cover, like npm publish, a write outside the project or a network call from Codex's sandbox. You only found out from a half-done report. Now:
Claude Code waits for you. wisp sends Claude's permission prompts to its own approve tool (--permission-prompt-tool). The task card shows Needs your approval: Allow Claude Code to run npm publish?, with Allow, Always allow in this task and Deny. The number keys pick an option. You can also type what to do instead: that denies the call and passes your words to Claude. Claude keeps its context, and the desktop notification says needs your approval. Always allow lets the same command or file through for the rest of the task without asking again.
Every other harness gets an Allow & retry card. When Codex, opencode, Grok, Cursor or Antigravity is refused, the report shows Blocked by permissions under its headline, listing each refused action. wisp reads these refusals from the harness itself: opencode's and Grok's stderr, Cursor's rejected tool calls, Antigravity's PERMISSION_DENIED steps and Claude's permission_denials. A new blocked block in the wisp protocol lets an agent report a refusal it can only see in a command's output, such as Codex's sandbox. Allow & retry sends a follow-up that runs those actions. For Claude and opencode only the exact actions are allowed (--allowedTools "Bash(npm publish)" or an opencode permission rule). Codex, Grok, Cursor and Antigravity can't allow a single action from the command line, so that one follow-up runs in Full auto, and the card says so. The task stays in Edits. Deny closes the card. API: POST /api/sessions/:id/approve {"turnId","allow"}. A live approval is answered like a question, through POST /api/sessions/:id/answer {"text":"Allow"}, wisp reply or the phone buttons.
Mention another task in a prompt. Right-click a task (or open its ⋯ menu) and pick Copy task ID, then paste the ID into the launcher, a follow-up box or the Rerun dialog. It shows as a chip with the task's title, like pasting an issue link in Linear, and stays a chip in the sent prompt, where clicking it opens that task. Backspace removes the whole chip and Cmd+Z brings it back. Pasting several IDs at once makes a chip for each. The agent gets the title and ID, and two new read-only tools sit next to ask_user: wisp_session reads a task (its title, state, harness, folder and branch, every request made in it, and a turn's report as markdown) and wisp_sessions lists your other tasks (filter by query, state, project, archived). Cursor and Antigravity runs can't take injected tools, so for them wisp puts each mentioned task's latest report in the prompt instead. Prompts store a mention as @[Title](wisp:<id>), so API and CLI callers can write one too. Task titles show @Title. The bridge routes (GET /api/internal/sessions/:id, GET /api/internal/session/:id?of=<id>) only take the per-run token.
Agents can answer questions about your wisp usage. Every task now gets a read-only wisp_stats tool next to ask_user. Ask a task "which tools do my tasks use?" or "what did Codex cost me this week?" and the agent gets a table of tool calls by kind (shell, read, edit, search, web, subagents, other MCP tools, questions), by harness, by project and by day. It also gets turns, cost, tokens and success rate, taken from the same numbers as Settings → Usage. Agents can narrow it with days (1–365, default 30), agent (a harness id) and project (a name or id, or current for the task's own project). Before, an agent had no tool for this and would either say it had no access or dig through ~/.wisp/sessions by hand. Kinds are wisp's categories, not each harness's own tool names. The bridge route (GET /api/internal/stats/:id) only takes the per-run token, like ask_user. Cursor and Antigravity runs can't take injected MCP tools, so they don't get it.
Changed
Mark done and the PR row are where you look. An opened task now has Mark done (or Reopen) in its top bar, next to Ship it, so you can close it out without scrolling to the composer. The PR row (checks, GitHub, Merge) moved from the top of the card to right above the composer, so after reading the result it's in view without scrolling back up.
Full auto is the default mode. New launches, API and CLI calls without a mode, scheduled and CI triggers, and recipes run with no permission prompts; tasks in git repos still get their own worktree and a checkpoint before every turn. If your saved launcher mode was Edits (the old default), wisp switches it to Full auto once; a Plan pick, or anything you choose afterwards, sticks. Issue triggers still start in Plan. The mode button now shows just an icon for Full auto and spells out Plan and Edits.
Switch harness or model in a follow-up. The follow-up box has a harness button, a small logo next to the mode button. It opens the same harness · model · thinking picker as the launcher. Pick another model and the follow-up resumes the same conversation on it. Pick another harness, say Codex on a task Claude Code started, and Codex starts its own conversation in the same folder. wisp hands it the turns it hasn't seen: each request, the report it got back and the files that were edited. Long histories keep the newest turns, up to about 24k characters. Each harness keeps its own conversation, so switching back to Claude Code resumes its session and only tells it what happened while it was away. Each turn's badge shows the harness that ran it. A queued follow-up keeps its own pick, so the run already going is not affected. On a session that Auto routed, switching counts as a miss for Auto's pick. API: POST /api/sessions/:id/reply {"text","agent","model","effort"}; CLI: wisp reply <id> "…" --agent codex --model …; MCP: wisp_reply takes agent, model and effort.
No harness yet? Install opencode in one click. When wisp finds no coding-agent CLI on this computer, the launcher shows one line, No coding agent here yet, with an Install opencode button. opencode's free models need no account, so a new user can go from npx wisp-agents to a running task in a few seconds. wisp installs it with npm into its own folder (~/.wisp/tools), so it needs no sudo and leaves your shell profile alone, and finds it there. Pressing Enter with no harness shakes that line and keeps your prompt; before, the launch failed with Auto found no installed harness. Settings → Agents has the same Install button on the opencode row. API: POST /api/agents/install {"id":"opencode"}.
wisp also finds opencode in ~/.opencode/bin, where its official install script puts it, when that folder isn't on PATH (common for the desktop app).
wisp Cloud preview page (cloud.html, linked from the site's nav and footer, and from a "Let it run while you sleep" section on the home page, between Also in the box and the install steps). It introduces a hosted wisp: managed remote machines, persistent queues that keep running while your laptop sleeps, and team dashboards. The local CLI stays free. You can try each part on the page: close the laptop and watch the cloud runs keep going, add machines, drag through one night of the queue, and filter a team dashboard by person. Signups for the waitlist go to PostHog.
One-line install: curl -fsSL https://mokh.xyz/install.sh | bash. The script (docs/install.sh, served by the site) checks for macOS or Linux, Node.js 20+ and npm, runs npm install -g wisp-agents (set WISP_VERSION to pin a release), and confirms that wisp is on your PATH. It prints a PATH fix when the npm bin folder isn't in it, and a link to npm's permissions guide when the global install is refused. The site and the README show it next to npx wisp-agents.
Homebrew:brew install m0xoo/tap/wisp-agents, from the new M0xoo/homebrew-tap repo (formula source in packaging/homebrew/wisp-agents.rb). It installs the CLI and browser mode; the formula removes the optional Electron package, which would otherwise download its ~300 MB binary into the Homebrew keg on first run.
Install tabs on the site. The hero command and step 2 of the install section both have tabs (npx, npm, curl, brew), like opencode's install block. Picking a tab in one switches the other, the copy button copies the command shown, and the note under it changes with the tab. Arrow keys, Home and End move between tabs. Type a task into the hero launcher and every tab hands back a command that installs wisp and starts that task (brew install m0xoo/tap/wisp-agents && wisp "…"). A command too long for the bar stays on one line and ends in …, and a long typed task no longer makes the page scroll sideways on phones.
Changed
The website moved to mokh.xyz. Canonical, Open Graph and structured-data URLs, sitemap.xml, robots.txt, llms.txt and the npm homepage now use the new domain. The old m0xoo.github.io/wisp/ address redirects there.
New website. The site opens on a working copy of the launcher: it types three tasks and sends them off, and their rows move from Working to Asking (answer it right there) to Done. Type your own task and press ↵ to get the npx wisp-agents "…" command that runs it for real. Further down it shows the real app instead of animated mock-ups. A scroll-driven "morning with wisp" walks through five real screens with a ticking clock: three tasks launched, a mid-run question with its desktop notification, a report, Ship it, and the merged PR. A Transcript / wisp report switch shows the same run both ways. Also new: a Compare section, a take-over table with each harness's resume command, and a terminal that replays real wisp CLI output. Screenshots follow your light or dark theme, and on phones the story becomes a simple timeline.
Fixed
Claude in Chrome now works in Claude Code tasks. Claude Code only turns the Chrome extension on by itself in interactive sessions; wisp runs it with -p, so its mcp__claude-in-chrome__* tools were missing. wisp now passes --chrome when you have Claude in Chrome enabled by default (/chrome in Claude Code, claudeInChromeDefaultEnabled in ~/.claude.json). Set WISP_CLAUDE_CHROME=1 or 0 to force it on or off.
An opencode task on a rate-limited model (often a free one) no longer sits at "Starting up" forever. opencode retries without printing anything, so wisp now reads its error log. If the rate limit hasn't cleared after 60s, wisp stops the task and says what to do: try again later, pick another model, or run opencode auth login.
wisp no longer crashes when the claude CLI quits before reading its prompt (a bad flag, an auth error, a crash on start). The task now ends with Claude's error message. Before, the unread prompt raised an unhandled EPIPE that took the whole server down.
config.json (projects, recipes, triggers, machines, settings) is now written atomically. Before, a crash or power loss mid-write could leave a cut-off file, which loaded as the defaults and was then saved over. Auto's learned state and the learned preferences are written the same way.
Idle API connections stay open longer than clients keep them, so a request sent on a connection the server was just closing no longer fails with ECONNRESET.
Refused requests (bad input, not found) no longer print a stack trace in the terminal running wisp.
Run / Re-run on a terminal, tests or actions block now shows the new output. Before, the command ran but the block kept its old output and an error toast said it failed.
Escape closes the menu or popover that is open, wherever focus is, and nothing else. Before, it could collapse the card behind a context menu and leave the menu open.
Settings switches, project edits, and deleting a recipe, trigger or subagent preset now show the server's error when it refuses the change, and leave things as they were. Before, nothing happened on screen; deleting a preset even said "Deleted".
Expanding cards no longer leaves a hidden file picker behind each time, and the launcher no longer adds another listener on every redraw.
openapi.json now documents subagent presets (/teams) and GET /uploads/{id}.
Internal
The server is split by area: server/index.js only wires things together. Routes live in server/routes/ behind a small route table (server/http/), and shared logic lives in server/services/ (launch, Ship it, Auto, uploads, phone, feedback). Helpers that were copied across files now come from server/util.js, and git always runs through one helper that honours WISP_GIT_BIN. Adapters are checked against their contract at startup, and claude.js uses the same spawn plumbing as the other harnesses. test/routes.test.js keeps the route table and openapi.json in sync. The HTTP API, CLI, SDK exports and the ~/.wisp files are unchanged.
The web UI's 5,300-line app.js is now about 40 modules under web/js/ (core/, ui/, features/), and app.css is 15 stylesheets that, concatenated in link order, are byte-for-byte the old file. A small event bus replaces the hooks that live updates used to reach open dialogs through, and one popover helper replaces about ten copies of open, outside-click and close code. No build step was added.
Calmer turn bar. The bar under each turn now fits on one line: the read / edited / ran / searched counts open the activity log themselves (no separate Activity link), and Restore to before this turn and Agent notes are icon buttons next to 👍 / 👎 that stay faint until you hover the turn. After a restore, Re-apply shows as text so the rolled-back state stays visible.
Fixed
Rating a turn with 👍 / 👎 no longer re-renders the whole card and replays its entrance animation. Every refetch re-inserted the turn nodes, which restarts their CSS animations; unchanged turns are now left in place.
Auto: wisp picks the harness and model for you, and it's now the default. Choose Auto in the harness picker (existing setups are switched to it once) and set a goal in the model slot: Balanced, Cheap or Best. For each task, Auto estimates how likely each harness · model you have installed is to do that kind of task (bug, UI, docs, tests…) well. It starts from a wisp-bench pilot that ships with wisp, then learns from your own outcomes: merged or closed PRs, 👍 / 👎, rollbacks, follow-ups and how the run ended. It weighs that against cost and time, and now and then tries an alternative so the estimates stay honest. Hover a card's harness logo to see why Auto picked it. Settings → Auto sets the goal, whether new models get tried, and which harnesses Auto may use; Open shows what it learned. A harness or model that shows up later (a new CLI, a new version) gets a few tasks that aren't marked important before Auto relies on it, and a new version starts from its family's record. Auto works for compare runs (a different pick each), for subagents (a team member set to Auto is picked on its first task), on machines (picked here from what the machine has, and that machine's outcomes feed the same router), and from the CLI, recipes, triggers and Telegram. Everything it learns stays in ~/.wisp/router.json. API: agent: "auto" on launch, GET/PATCH /api/router, POST /api/router/reset.
Fixed
Claude Code tasks that say "I'll report once it has finished" now do. When Claude ends its turn with background work still running (a test run, a build, a background subagent), the task stays in Working and shows Waiting on background task: …. When the work finishes, Claude's follow-up report becomes the task's answer. Before, wisp stopped Claude 3 seconds after its first reply, which also killed the background work, and the task ended on that "I'll report…" line. Stop and follow-ups still work while it waits, and it waits at most an hour. A reply that is already a full answer, for example one that leaves a dev server running, still ends the task right away. All harnesses are now also told to wait for their background work before finishing.
Follow-ups no longer end on "The agent finished without a message." When a Claude Code run left background commands behind, the next follow-up ended right away with an empty answer, and the agent was cut off before it started on your message, so you had to send it again. wisp now waits for the answer to your follow-up.
Start tasks from Telegram. In Phone → Telegram, switch on Start tasks from Telegram, then send your bot /new fix the flaky login test. The task runs in your default project (the last used, or the one you pick in that section) with the launcher's harness, model and mode. /new @api add rate limiting picks a project by name, and /projects lists them. The bot replies ▶ Started, and replying to that message sends a follow-up. Its card shows Telegram as where it came from. It's off by default, only the linked chat can start tasks, and a /new sent while wisp was offline is not run later. API: PUT /api/remote with telegram.allowStart and telegram.project.
Feedback for the wisp team. Each finished turn has a small 👍 / 👎 at the end of its trail. It stays faint until you hover the turn. Click again to take the rating back. 👎 opens one optional line for what went wrong (↵ sends, Esc skips). You can also type /feedback <note> in the launcher or any follow-up box, or use ⌘K → Send feedback… for a short dialog. Bare /feedback opens the dialog too. It thanks you right away; offline, the feedback is kept and sent later. It goes to PostHog with the harness, model and status of the run, never your prompt, output or code. It's sent even with usage stats off, and WISP_FEEDBACK=0 turns it off. API: POST /api/feedback.
Changed
Switch modes in a follow-up. Each task's follow-up box now has the same mode button as the launcher (Plan, Edits, Full auto; Shift+Tab cycles). Start in Plan, read the plan, then pick Edits and say "go ahead": that follow-up and the ones after it run in the new mode. Any mode can switch to any other. The pick only takes effect when you send, and the follow-up shows the new mode next to its model. A follow-up queued while the agent works keeps its own mode, so the run already going isn't affected. A task started in Plan has no worktree, so edits after switching land in the project folder. API: mode on POST /api/sessions/{id}/reply; CLI: wisp reply <id> "…" --mode edits; MCP: mode on wisp_reply.
Compare more than two harnesses. Clicking vs in the launcher adds a second harness as before, and a + at the end of the pill now adds a third and a fourth (four side by side, the server's limit). Each one sits behind its own vs: hover it for × to drop just that harness. A new one starts on a harness that isn't in the pill yet. With three or four, the pill shows logos only (hover a logo for harness, model and thinking level), and the mode and Subagents labels fold to icons so the bar stays on one line.
Fixed
Changing a Telegram setting through the API without a token no longer clears the bot token and unlinks the chat. After a settings change, Telegram no longer re-delivers messages wisp already handled.
Following up on a task now keeps it on screen as it moves to Working, and again when it lands back in Needs you. On a long card, the previous answer folding away used to pull the whole card off the top of the screen, so you lost the task you had just replied to. The composer now holds still and the new exchange appears where you were typing.
Settings → Agents → Re-check looks for harnesses and models again, on this computer and on every online machine. Install a CLI, sign in or add a provider (say a new opencode provider), come back to wisp and click it: the list, the model counts and the launcher's harness picker update without restarting wisp, and a toast says what changed ("opencode +12 models", "Codex found"). Each harness now shows how many models it has, and one that isn't installed says to install its CLI and sign in. Before, wisp kept the first detection until it restarted. API: POST /api/agents/recheck, or GET /api/agents?fresh=1 for this computer only.
Changed
The calm feed (nothing needs you) looks tidier. Nothing needs you gets a green dot and lines up with the Done and Archived headings, and folded groups now sit close together instead of a full section gap apart. Folded headings highlight on hover so they read as buttons.
Fixed
In Settings → Machines, the add/edit machine form lays out properly again: labels sit left above their fields, and the SSH host and Name inputs are full height instead of squashed and centred. The Settings row style was leaking into it.
Machines shows Update to X next to the version when a connected machine runs an older wisp than this computer. One click restarts wisp there as this version (you're asked first if agents are working on it). Before, reconnecting kept the old wisp running, and you had to find Restart wisp there in the ⋯ menu. Machines with a custom start command don't show it, because that command decides the version. Machines carry it as outdated: "<version>".
Changed
The app responds right away instead of waiting for the server. A new task clears the prompt box and shows up in Working as a "Starting up" card while wisp makes its worktree. A follow-up or an answer clears the composer and moves the card to Working with your message as the live turn. Mark done, Reopen, Archive, Delete, Stop, Rerun, recipe runs and Keep this in Compare also act immediately. If the server refuses, the change is undone, your text and attachments come back, and an error toast says why.
The web app uses the same icon as the site and the desktop app (the glowing orb on a dark tile) for its browser tab, home-screen shortcut and notifications.
The launcher's Run on button sits right after the attach button, so the bar reads attachments, where it runs, then mode, harness, subagents and worktree.
Fixed
The launcher's harness picker now lists the harnesses installed on the places picked under Run on, not the ones on this computer. Any picked place that has a harness makes it pickable (a union), with its models taken from a place that has it; a harness missing on some of them says Not on <machine>, and launching names the place it's missing on. Before, running only on a machine still offered Claude when only this computer had it, and hid a harness only the machine had. Changing the places (or a machine coming online) moves the pick to an installed harness if the current one isn't there. Machines now report each harness's models, thinking levels and whether it can lead subagents.
Machines finds hosts for you. Found on this computer lists the hosts this computer already reaches: Host entries in ~/.ssh/config, your Tailscale peers (tailscale status --json, with an online dot; phones and tablets are skipped) and hosts in ~/.ssh/known_hosts (code hosts such as github.com, hashed entries and non-default ports are skipped). Hosts you've already added aren't listed. Click one to fill in the form, then press Connect. The SSH host field suggests the same hosts. API: GET /api/machines/discover.
Changed
Ship it only shows when there is something to ship. It used to sit on every finished worktree task, even one that changed nothing or whose work was already on its PR. Now it appears (on the card bar, next to the composer, in the right-click menu, in ⌘K and on S) only if the worktree has uncommitted changes or commits its PR doesn't have yet. Push to PR #N shows only when there are new changes to push. wisp checks the worktree when a turn ends, after Ship it, a revert, a rollback or an action command, and at startup. Summaries carry it as toShip: { dirty, ahead, unpushed }.
Fixed
Updating from the top-bar Update button now restarts wisp on the port it was running on. Before, a wisp started with --port came back on 4477 and the open tab said "wisp did not come back".
Machines: run agents on your other computers (a dev box, a GPU server, a cloud VM) over SSH. Machines in the top bar (also Settings → Machines, ⌘K) takes an SSH host (suggested from ~/.ssh/config) and connects: wisp signs in with your keys, finds Node.js there, starts npx -y wisp-agents@<this version> detached (or reuses the wisp already running), and opens an ssh -L tunnel back. Each step shows live, and a failure says what to run (ssh-copy-id, install Node, restart an older wisp). Then pick projects on the machine (its git repos are listed, folders autocomplete). The launcher's Run on button (next to the harness) picks where a prompt goes: this computer, a machine, or several at once. Each place gets its own row of project chips, and one Enter launches on all of them in parallel, as one task card with a row per machine (its own project, badge and status) and a single "ready on 3 machines" toast. A machine picked while disconnected connects. Each run executes the agent on its machine, with worktrees, diffs, rollback, ask_user, subagents, compare and Ship it working as usual. Its sessions join your feed with a machine badge, answer questions from your phone like any other, open in VS Code / Cursor / Windsurf over Remote - SSH, and Take over opens ssh -t in your terminal. Nothing new listens on the network: wisp over there stays on its own loopback, keeps running (agents too) when this computer sleeps, and wisp reconnects when it wakes. API: /api/machines, machine and groupId (to group runs launched in several places) on POST /sessions, machine on /projects; remote sessions are <id>@<machine> everywhere (CLI, SDK, MCP).
Grok Build harness (grok, --agent grok). wisp runs grok --output-format streaming-messages-json headless. It lists the models from grok models with each model's thinking levels (Low to X-High), maps Plan / Edits / Full auto to grok's permission modes, and resumes follow-ups with --resume. Grok can lead a team of subagents, and it asks you questions through wisp's ask_user. The tool reaches grok through a throwaway --agent definition that keeps grok's stock prompt. Diffs come from search_replace and write. Cost shows in Usage when grok reports it.
Changed
The launcher's Recipes row no longer has a Manage link. Recipes are still one click away from the Automations button in the top bar.
Every text field, textarea, select and checkbox now shares one look from a small form kit in web/css/app.css (.input, .textarea, .field, .check, plus .mono / .sm / .lg): the same border, radius, height, placeholder colour, hover and focus ring across Settings, Phone, Automations, Ship it, Run recipe, Projects, the model picker search, the Subagents editor and question cards. Checkboxes use the accent colour.
Fixed
A lead with subagents now uses them instead of its harness's built-in subagents. Claude Code leads run with Task, Agent and Workflow disallowed, opencode leads with permission.task: "deny", and every lead is told to use delegate in place of its own subagent tool. Without a team, nothing changes.
The selects in the recipe and trigger forms and in the Ship it / Run dialogs had lost their dropdown chevron.
Settings and Phone text fields, and every select, showed no focus state.
Subagents: give the lead agent a team on other harnesses, e.g. a Claude Code run that hands tests to Codex and research to opencode. The launcher's Subagents button opens an editor: up to 6 members, each with its own harness, model and thinking level, a name, a Use for note ("image and UI work", "research") that tells the lead when to pick it, and a ×N copies stepper: Codex ×3 gives the lead three identical Codex agents (codex#1…codex#3, 12 agents in all). Several tasks sent to codex in one call go to free copies and run side by side; naming a copy continues its conversation. Save as preset stores the team; Presets loads or deletes one. The lead gets a delegate tool on wisp's injected MCP server (Claude Code, Codex, opencode; Cursor and Antigravity can be subagents but not leads). Tasks in one call run in parallel in the lead's folder, and delegating to the same member again resumes its conversation. Feed cards show the lead's logo with its subagents' logos stacked behind it (they fan out on hover and pulse while working). A running card lists each subagent's live status, and a finished turn shows one report card per delegated task that opens into that subagent's own presentation. Subagent edits count as the lead's, so they appear in its files, diffs, rollback and Ship it. API: subagents / preset on POST /sessions, /teams for presets, PATCH /sessions/:id { subagents }; wisp_launch takes them too.
Actionable presentation blocks: a new actions block (items: [{label, command, detail, confirm} | {label, reply}]) renders 1-click command buttons that run in the session's project folder (POST /api/sessions/:id/exec), show live exit code, duration and output inline, and offer a Send failure to agent follow-up button when a command fails. terminal and tests blocks with a command now have a Run / Re-run button, and diff and files blocks have a per-file Revert / Re-apply button (POST /api/sessions/:id/revert).
Session checkpoints & one-click rollback: before each turn starts, wisp snapshots the git working tree (git stash create || git rev-parse --verify HEAD, pinned at refs/wisp/checkpoints/<session>/<turn>) along with pre-turn untracked files and per-file before snapshots. Each turn's trail bar, session menu, and command palette include Restore to before this turn (and Re-apply turn / Undo) via POST /api/sessions/:id/rollback.
Worktree per task: in a git repo, each launch runs in its own git worktree on a wisp/<task>-<id> branch made from origin/main (freshly fetched), under ~/.wisp/worktrees/. Parallel agents never edit the same checkout, and the agent is told where it is. On by default; toggle Worktree / In place in the launcher, Settings → Worktree per task, --no-worktree on the CLI or worktree: false in the API. Plan mode and folders that aren't repos run in place. Archiving removes a clean worktree (the branch stays, a follow-up brings it back); deleting the session removes it, and its branch unless it has commits.
Ship it (S, the card's Ship it button, or wisp ship <id>): commits the worktree with a conventional message drafted from the presentation, pushes the branch and opens a PR with gh pr create (optionally as a draft). The card then shows the PR: CI checks, review state and merges, polled every minute; it comes back to Needs you when checks finish, a review lands, or it merges. Merge (wisp merge <id>) marks a draft ready, squash-merges with --match-head-commit, fast-forwards your local base branch and removes the worktree.
Compare: the vs button in the launcher (or --compare codex,opencode:model on the CLI, compare: [...] in the API) runs the same task on a second harness, each in its own worktree. Compare opens the runs side by side: time, cost, files and lines changed, tests, questions, and each presentation. Keep this archives the others and deletes their worktrees.
Recipes: saved prompts where {name} becomes a field to fill in, e.g. Prepare release {version}…. Create them from ⌘K → Recipes or Save as recipe… on any card; run them from ⌘K → Run: …, or wisp run release 0.8.0 / wisp recipes. A recipe can pin a project, harness, model and mode.
Triggers start sessions without you, while wisp is open: a cron schedule (local time), CI failed (a GitHub Actions run fails; polled with gh run list), or issue labelled (polled with gh issue list). Each runs a prompt or a recipe, with the run's or issue's details filled in. The first check only records what already exists. Issue triggers default to Plan mode, since issue text is untrusted. ⌘K → Triggers.
Usage view (⌘K → Usage, Settings → Usage): spend per day, and spend, turns, success rate, questions and average time by project and by harness · model, over 7, 30 or 90 days. Harnesses that don't report cost are counted but not priced. wisp usage, GET /api/usage.
Phone (Settings → Phone): questions and finished tasks go to ntfy and/or a Telegram bot. Tap an option to answer; on Telegram, reply to a message to answer in your own words or follow up, and tap Mark done. wisp polls both, so nothing has to reach your machine. ntfy answers are HMAC-signed; Telegram is linked to one chat with a one-time code.
API: POST /sessions takes worktree and compare; new /sessions/:id/ship, /pr, /merge, /recipes, /triggers, /remote, /usage (all in GET /api/openapi.json). SDK: ship, merge, recipes, saveRecipe, runRecipe, triggers, usage. Summaries carry branch, pr, groupId and origin.
A graph presentation block for explaining merges and conflicts: a PR card with the branches as coloured lanes (newest commit first), the merge target on top (conflicts / ready / merged), branch tags on each tip, and the base commits that conflict highlighted and numbered. Under it: each conflict with file, ours vs theirs and how it was resolved (hover one to light up the commit it came from), checks as chips, and an optional action button (e.g. Resolve conflicts) that replies to the agent. Plain-markdown rendering for the CLI, MCP and chat apps too.
Back to top from the keyboard: U (or Home) scrolls the feed to the top. The back-to-top button that appears once you scroll shows the U key next to its arrow, and the shortcut is listed under ?.
Changed
The bar under an open task's title only says how the run was set up: the mode when it's Plan or Full auto, the branch, and what launched it. Agent, model and path, which the card header already shows, moved into its tooltip. Compare stays in the task header and Done next to the composer; Take over is an icon. Ship it now also sits next to Mark done under the composer, so you can ship right after reading the result.
Memory, Automations, Usage and Phone open inside Settings when you reach them from there: they take the main area next to the sidebar, with a Back button (or Esc) that returns to the section you came from. Only the X closes Settings. From the top bar or ⌘K they still open on their own.
Settings is a sidebar and a pane instead of one long list: General (editor, terminal, worktree per task), Notifications (desktop, phone), Agents (each harness with version or error; a red dot in the sidebar when one is missing), Memory, Automations (recipes and triggers, each with Manage), Usage (last 7 days at a glance) and Privacy. It reopens on the last section; ↑/↓ move between sections. Phones get the sections as a scrolling tab strip.
Recipes and triggers are out in the open instead of only behind Settings and ⌘K: an Automations button in the top bar (badge = triggers on, tooltip = what fires next), and a Recipes row under the launcher with one-click chips for the picked project's recipes (and those tied to none), Save as recipe for whatever is in the box, and Manage. Each recipe in Automations has a Schedule button that starts a schedule trigger running it.
The launcher stays reachable while you scroll: once it passes under the top bar it docks there as a one-line bar (project chip + draft). Click it, or press ⌘L / N, and the full launcher opens in place over the feed, without scrolling you back to the top. Esc or clicking away folds it again. Cards selected with J / K or opened scroll into view below it.
The launcher's vs (compare) control is part of the harness pill instead of a separate button. Turned on, the second harness joins the same pill (Claude Code ⌄ vs opencode ⌄); hover vs for × to stop comparing. The bar no longer wraps onto a second line: on narrow widths model names, then labels, fold away and the logos stay.
A compare launch is one task in the feed instead of a card per harness: the prompt on top, a row per run (harness, model, live status; each expands in place), and Compare for the side-by-side view. The task sits in Needs you while any run asks, Working while any runs, and Done once all are; ✓ marks every finished run done. Compare runs keep the prompt as their title. Keep this turns the kept run back into an ordinary card. See docs/feed.md.
Fixed
Clicking a recipe no longer runs it straight away when it has no fields to fill in. Its dialog now always stays open until you press Run (the button has focus, so Enter works too).
flow blocks that don't fit across the card wrap onto rows that alternate direction (left to right, then right to left), joined by a short line, instead of squeezing every step into a narrow column. Loop-back lines go around the cards instead of crossing them. Phone width still stacks the steps.
The launcher no longer flashes its focus ring when you open the harness pickers or switch between them (e.g. opencode, then Claude Code in a compare). The ring stays on while a picker or the mode menu is open; the rerun dialog behaves the same way.
Launching a compare no longer freezes wisp for a few seconds. The runs share one git fetch and their worktrees are checked out in parallel. A fresh worktree's first checkpoint is its base commit, so it no longer runs blocking git calls that stalled the server and the live feed. With fake harnesses, a 2-way launch went from 1.5 s to 0.7 s and a 3-way launch from 2.3 s to 0.9 s. The send button shows a spinner while the launch sets up. If one worktree fails, the ones already created are removed.
Scrolling in the feed stays put. When a card changes section in the background (a task finishes, asks or is removed), the card you're reading no longer jumps; one whose edge just peeks onto the screen no longer drags the view with it. Mark done, Archive and Delete put the next card where the old one was and focus it, so E / A / j triage continues from the same spot. j / k show the header of tall cards, switching project filters returns to the top of the feed, and a task launched from the box scrolls into view when its card lands off-screen. The rules are written down in docs/feed.md.
Update button next to Search when a newer wisp-agents is on npm (checked at start, then once a day while wisp stays open — survives sleep, and the UI re-asks on focus). One click runs npm install -g wisp-agents@latest (or relaunches through npx …@latest) and restarts wisp; it asks first if agents are running. Git checkouts just get the git pull hint. Opt out with WISP_UPDATE_CHECK=0.
Rerun a task: right-click a card → Rerun… (or ⇧R, or ⌘K → Rerun). A dialog opens with the original prompt (editable), attachments and mode, plus the harness / model / thinking-level picker. You can pick a different setup without changing the launcher's defaults. Rerunning starts a new session in the same folder and keeps the original.
Learning from sessions: after a session goes quiet (90 s, or right away when marked done or archived), a background retro (same harness, read-only, no tools; Haiku on Claude Code) records your preferences and the agent's lessons (a mistake and the way that worked), globally and per project. Later turns get them in their system prompt. Stored in ~/.wisp/preferences.json; API GET/PUT /api/preferences. See Learning.
Memory view (Settings → Memory → Open, or ⌘K → Memory): a tab per scope with preferences and lessons, the session that taught each item, a token estimate, and inline edit / add / forget. The Retros tab is a timeline of every retro run (last 30), including nothing new and failures with the reason, plus a live entry while one runs. Switching Memory off stops both learning and applying.
Fixed
The card you were reading stays in view after collapsing a section, and the open card is followed when it moves to another section.
Antigravity CLI adapter (agy --input-format stream-json --output-format stream-json). Plan → --mode plan, Edits → --mode accept-edits, Full auto → --dangerously-skip-permissions; models from agy models, --effort low|medium|high, follow-ups via --conversation. Like Cursor it has no flag to inject MCP servers, so Antigravity agents ask with a question block. Override the binary with WISP_ANTIGRAVITY_BIN.
Integrations: wisp can now be driven by other agents, editors and scripts. Reference: docs/integrations.md.
wisp mcp: a stdio MCP server with wisp_launch, wisp_wait, wisp_get, wisp_reply, wisp_list, wisp_stop and wisp_agents, so Claude Code, Codex, opencode, Cursor or Claude Desktop can hand tasks to wisp and read the results.
Scripting CLI: wisp ls, show, wait, reply, stop, plus --json, --wait, --cwd, --model and --effort on launches. Use wisp -- … for a task whose first word is a command name.
JS SDK: import { WispClient } from 'wisp-agents'. It connects to the running wisp (or starts a headless one) and covers launch, wait, reply, presentation, uploads and the event stream. wisp-agents/render, /protocol and /server are exported too.
HTTP API: GET /api/health, GET /api/openapi.json (OpenAPI 3.1), GET /api/sessions/:id/wait (long-poll until it finishes or asks), GET /api/sessions/:id/presentation (blocks + markdown, ?format=markdown), and state/agent/project/cwd/archived/limit filters on GET /api/sessions.
~/.wisp/server.json advertises the running server's URL, port and pid.
Webhooks (PATCH /api/config {"webhooks": [...]}): POST session.finished / session.waiting with the presentation as markdown.
Fixed
Clicking a harness in the harness · model picker (or Enter on it) switches to it with the model and thinking level shown, instead of only previewing its models.
Antigravity models are listed with newer agy builds that print tab-separated agy models output.
A harness CLI installed while wisp is running is detected without restarting wisp.
Anonymous usage stats and error reports via PostHog: which agents, models and modes are used, run outcomes and timings, and errors (agent failures, server errors, crashes) with paths scrubbed. Never sends prompts, output, paths or project names. Turn it off in Settings → Anonymous usage stats, or with WISP_TELEMETRY=0 / DO_NOT_TRACK=1; always off when CI is set. See the Telemetry section of the README.
Thinking level in the harness · model picker: Default / Low / Medium / High / X-High / Max under the model list, remembered per harness and shown on the launcher button. Claude Code gets --effort, Codex -c model_reasoning_effort=…, opencode --variant (levels read per model from opencode models --verbose). Cursor bakes effort into its model ids, so it has no separate selector.
Take over in cmux: opens a new cmux workspace in the project running the resume command. Auto-detected when wisp is started from a cmux terminal.
Codex (codex exec --json), opencode (opencode run --format json) and Cursor CLI (cursor-agent -p) adapters, next to Claude Code. Pick the harness and model together in the launcher, or pass --agent on the CLI.
One harness · model picker in the launcher: harnesses (with logos) on the left, that harness's models on the right, searchable and grouped by provider for opencode. The model is remembered per harness. opencode and Cursor model lists are read from the CLIs themselves.
ask_user works in Codex and opencode (injected MCP server). Cursor has no flag to inject one, so Cursor agents ask with a question block instead.
Diffs for harnesses that only report "file changed" are taken against the file as it was when the turn started (git stash create), or the agent's own before/after content when available.
Windows: harness CLIs installed through npm (.cmd shims) and PowerShell-only installs now start. npm shims run as node <script>, so arguments reach the CLI unchanged. Stopping a run ends the whole process tree (taskkill /T).
Take over on Windows (Windows Terminal, PowerShell, Command Prompt) and on more Linux terminals (GNOME Terminal, Konsole, Xfce Terminal, Alacritty, xterm). The settings list only shows terminals for your OS.
Test suite (npm test) and CI on Linux, macOS, and Windows, plus npm run test:real to smoke-test the real CLIs.
Images an agent shows are copied into ~/.wisp/shots/<session>/ when the turn finishes, so screenshots saved to /tmp survive reboots. Existing sessions are backfilled on startup if their files still exist. Deleting a session deletes its copies.
Command palette (⌘K / Ctrl+K, or the Search button in the top bar): fuzzy search across sessions (title, project, summary, agent), project filters, global actions, and actions on the focused session.
Archive tasks: right-click a task (or press A) to archive it. Archived tasks move to a collapsed Archived section and keep their state; a new question or turn brings them back.
Right-click menus on tasks (open, rename, mark done, archive, delete), project chips (remove) and queued follow-ups (cancel).
⌘L / Ctrl+L focuses the launcher from anywhere. M renames the focused item. Menus show key hints.
Queued follow-ups are listed above the composer and can be cancelled from an × button, with confirmation.
Task cards and requests show the harness logo, with a "harness · model" tooltip.
Fixed
Image blocks pointing outside the project (e.g. /tmp/shot.png) no longer render broken.
Replying to a finished agent no longer gets queued behind a CLI that keeps running after its result. A run now ends on exit, 3s after its result, or when a follow-up arrives.
Removing a project or deleting a session no longer fails silently (body-less DELETEs now send a JSON content type).
The notifications setting is a real on/off switch.
Finished cards fold up unless you are looking at them.
Shortcut labels read ⌘L on macOS and Ctrl+L elsewhere, and modifier shortcuts work on non-Latin keyboard layouts.
Changed
A follow-up sent while an agent is working now goes out right away: wisp interrupts the run and resumes the session with your message. To get the old behaviour back (wait until the run finishes), turn on the Queue toggle next to the send button. It shows while the agent is working, and wisp remembers the choice.
The Plan / Edits / Full auto control is now one compact button with a menu. It shows a label only for a non-default mode; Shift+Tab in the prompt cycles modes.
The top bar no longer shows harness status; Settings lists each harness with its version.
Desktop app: sending a message no longer opens an extra wisp icon in the Dock. The ask_user helper now runs the Electron binary as plain Node (ELECTRON_RUN_AS_NODE=1) instead of starting a second app.
Desktop app: native right-click menu with spelling suggestions, Copy Link, undo/redo, cut/copy/paste and select all. Inspect Element appears only in dev mode.
Fixed
macOS: the Dock shows "wisp" with the wisp icon instead of "Electron". The app now launches from a renamed clone of Electron.app, cached per Electron version under ~/Library/Caches/wisp.
npx wisp-agents opens the desktop app. Electron is now an optional dependency that is fetched on first launch. The CLI starts the app detached and falls back to the browser when Electron isn't available.
--browser flag to keep the browser-only behaviour.
Fixed
Parsing of the wisp fence is more robust: malformed blocks are salvaged instead of dropped, and the raw turn text is kept.