Claude DevFleet Multi-Agent Work

Community skill. Full credit goes to the original Claude DevFleet skill author.

When to Use

Use this skill when a coding job has parts that agents can do at the same time.

Each agent works in its own Git worktree. This keeps file changes apart until DevFleet merges them.

Do not use this skill for a small task that one agent can finish fast. Do not split work when two tasks must edit the same files at the same time.

Setup

Claude DevFleet must be running. Add its MCP server:

claude mcp add devfleet --transport http http://localhost:18801/mcp

If the server cannot be reached, stop and tell the user how to start or connect it. Do not make up tool results.

Tools

Work Steps

1. Call plan_project(prompt="...").
2. Check that each mission has one clear goal.
3. Check all mission links. A mission must not depend on itself. The links must not form a loop.
4. Show the plan to the user. Ask for approval before agents change code.
5. Start missions that have no unmet needs.
6. Use get_mission_status() to check work. Do not poll in a tight loop.
7. Let missions with auto_dispatch=true start after their needs pass.
8. When a mission ends, call get_report().
9. Check its changed files, completed work, tests, errors, and next steps.
10. Give the user one clear final report.

Safety Rules

• Use the right repo path.
• Keep each mission small and clear.
• Give two agents separate files or areas when you can.
• Do not run two missions that may edit the same file at once.
• Do not dispatch a mission twice.
• Do not claim a merge or test passed unless the report says it did.
• If a mission fails, read its report before retrying.
• If a mission times out, check its status. It may still be running.
• Cancel work only when the user asks, the task is unsafe, or the agent is stuck.
• If an auto merge fails, report the clash. Do not hide it.
• If one mission fails, do not start missions that need its work.
• If the plan changes, show the new plan before sending more agents.
• Do not add analytics, telemetry, web calls, or other outside calls.

Concrete Example

User request:

Build a REST API with sign-in and tests.

Call:

plan_project(
  prompt="Build a REST API with sign-in and tests. Keep API code, sign-in code, and tests in separate missions when safe."
)

A good plan may be:

M1: Add the base API routes
M2: Add sign-in
M3: Add tests, depends_on=[M1, M2]

Show this plan and get approval. Then start the first two missions:

dispatch_mission(mission_id="M1")
dispatch_mission(mission_id="M2")

Set M3 to start after both needs pass:

create_mission(
  project_id="PROJECT_ID",
  title="Add API and sign-in tests",
  prompt="Add tests for the API routes and sign-in flow. Run the test suite.",
  depends_on=["M1", "M2"],
  auto_dispatch=true
)

Check progress:

get_mission_status(mission_id="M1")
get_mission_status(mission_id="M2")

When each mission ends, read its report:

get_report(mission_id="M1")
get_report(mission_id="M2")
get_report(mission_id="M3")

Tell the user what changed, which tests ran, what failed, and what still needs work.