Guide
Multi-agent coding across Claude Code, Codex and more
In wisp, one lead agent can run a team of coding agents on other harnesses: Claude Code plans the change and hands the tests to three Codex copies, which work in parallel. The lead gets a delegate tool over MCP, and every member’s edits land in its worktree, its diff and its pull request.
Free and open source · Claude Code, Codex, Cursor CLI, opencode, Antigravity CLI, Grok Build · macOS, Linux, Windows (beta) · Node 20+
Four ways to run agents together
| Pattern | When | In wisp |
|---|---|---|
| Parallel tasks | Independent tasks, one per agent | Launch them back to back; a worktree each. Guide |
| Compare | One task, the best of several attempts | vs in the launcher, --compare. Guide |
| Subagents | One big task, split by a lead | Subagents in the launcher (below) |
| Machines | More agents than one computer handles | Run on in the launcher. Guide |
They combine: a lead with a team, running on a dev box, compared against a second lead.
Give the lead a team
The launcher’s Subagents button sets up a team of up to 6 members. Each has a harness, a model and thinking level, a name, a Use for note that tells the lead when to pick it, and a copies count:
| Member | Harness | Use for |
|---|---|---|
| lead | Claude Code | Plans the change, reviews, ships |
codex#1…codex#3 | Codex ×3 | Writing and fixing tests, one module each |
docs | opencode, a cheap model | README and changelog |
Claude Code, Codex, opencode and Grok Build can lead; Cursor CLI and Antigravity CLI can be members too. Save as preset keeps a team for later.
How delegation works
- The lead gets a
delegatetool on wisp’s MCP server. Tasks it sends in one call run in parallel, in the lead’s folder. - Delegating to the same member again resumes that member’s conversation, so it keeps its context.
- Members’ edits count as the lead’s: they show up in its diff, its checkpoints and rollback, and its pull request.
- The call returns each member’s report and the files it changed, and wisp shows those reports next to the lead’s. Members don’t ask you questions; they make the call and note the assumption, and the lead decides.
From the API, pass subagents or a saved preset to POST /sessions. See integrations.
Orchestrate from the agent you already use
wisp ships an MCP server. Add wisp mcp to Claude Code, Codex, Cursor or opencode, and the agent in your editor can launch tasks into wisp’s queue (wisp_launch), wait for them (wisp_wait) and read the reports back, on any harness and any machine.
claude mcp add wisp -- npx -y wisp-agents mcp
# or install the agent skill
npx wisp-agents skill install --agent allQuestions
- Do subagents work in the same worktree?
- Yes. Members work in the lead’s folder, so their edits are part of one branch, one diff and one pull request. Separate tasks get separate worktrees.
- Can the team run on another machine?
- Yes. When a task runs on a machine, the lead and its team run there, with the harnesses that machine has.
- Which agents can lead?
- Claude Code, Codex, opencode and Grok Build can lead a team. Cursor CLI and Antigravity CLI can be members.
Related guides
- Run Claude Code and Codex in parallelSame task on both agents, or different tasks at once, each on its own branch.
- Run coding agents on multiple machinesYour desktop, a dev box, a cloud VM: one prompt, several machines, one queue.
- Claude Code vs Codex, on your own repoRun the same task on both and compare tests, time, cost and the diff.