Skip to content

MCP from your harness

You don’t have to leave the agent you already use. wisp mcp serves wisp’s tools over MCP, so Claude Code, Codex, opencode or any MCP client can hand work to wisp: start a task in the background, wait for it, read its report and follow up.

Add wisp to your harness

One command, in your repo, does it for Claude Code and Codex:

Terminal window
wisp skill install --agent claude # .mcp.json, a SessionEnd hook, .claude/skills/wisp
wisp skill install --agent codex # [mcp_servers.wisp] in ~/.codex/config.toml
wisp skill install --agent all # Claude Code, Codex and opencode
  • Claude Code: .mcp.json gets wisp’s MCP server (wisp mcp), .claude/settings.json a SessionEnd hook (each session you end is brought into wisp with wisp adopt --hook), and .claude/skills/wisp/SKILL.md tells it when to use wisp. Commit them and your team has it too.
  • Codex: ~/.codex/config.toml gets [mcp_servers.wisp]. Codex has no end-of-session hook: bring its sessions in with wisp adopt --all --since 1d.
  • opencode: the skill goes in place; add the server by hand in opencode.json:
{ "mcp": { "wisp": { "type": "local", "command": ["wisp", "mcp"] } } }

Cursor (.cursor/mcp.json) and any other MCP client: point it at wisp mcp on stdio.

{ "mcpServers": { "wisp": { "command": "wisp", "args": ["mcp"] } } }

No global wisp? Use npx -y wisp-agents mcp in its place. Tasks start in the folder the client started wisp mcp in, usually your project.

The skill on its own

The wisp skill tells an agent when to hand work to wisp, how to brief it, how to wait, and to pass a waiting agent’s question on to you instead of answering it.

Terminal window
wisp skill # print it, to read first
wisp skill install --agent codex --project # .codex/skills/wisp in this repo

--agent takes claude, codex, opencode or all. --project writes into the repo, to commit for your team.

The tools

Tool What it does
wisp_launch Start a task on any harness, model, effort, mode or machine. Returns its id at once; wait blocks.
wisp_list Tasks, newest first, with their state and a one-line summary.
wisp_get A task’s state and latest report as markdown, and any question it is waiting on.
wisp_wait Block until a task finishes or asks something.
wisp_reply Follow up, or answer its question. Can switch harness, model, effort or mode.
wisp_stop Stop a running task.
wisp_agents The harnesses installed, with their models and thinking levels, and your machines.

An agent running inside a wisp task gets more: ask_user, wisp_pr, wisp_watch, the crew and connectors, and resume_later. Every tool is in MCP tools.

What to ask your agent

  • “Send the flaky-test fix to wisp on Codex and keep going on the UI.”
  • “Start three wisp tasks, one per failing package, and tell me when they’re done.”
  • “Ask wisp what my last task changed.”
  • “Start a wisp task to fix the login bug, with the goal the e2e suite passes.” (See Goals.)

Follow what it started

Tasks your agent starts are ordinary wisp tasks.

They show in your task list like any other. Open one to read it, answer it or follow up. J and K move between them.

In the browser (wisp --browser, or your cloud wisp) it’s the same window, menus and keys.

Sessions you run outside wisp

Work you did in your harness without wisp can come in too, so it’s in your history:

Terminal window
wisp adopt --all --since 7d