Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@claude-operator assign a task to fix the login bug in project alpha"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
TypeScript MCP server that turns any Claude Code session into an autonomous operator — decomposes goals, spawns and coordinates worker sessions across multiple projects, manages persistent two-tier memory, enforces guardrails, and learns from retrospective human review.
npm install -g claude-operator
claude-operator init
claude-operator startinit scaffolds ~/.claude-operator/ (SQLite DB, bus dir, config).start launches the MCP server on stdio + dashboard at http://localhost:7373.
Related MCP server: claude-swarm
Add to your ~/.claude/settings.json:
{
"mcpServers": {
"claude-operator": {
"command": "claude-operator",
"args": ["start", "--mcp"]
}
}
}Or point at the local dev build:
{
"mcpServers": {
"claude-operator": {
"command": "node",
"args": ["/path/to/claude-operator/packages/server/dist/server.js"]
}
}
}Operator session (Claude Code)
│ MCP tools (10)
▼
┌─────────────────────────────────┐
│ claude-operator MCP server │
│ ┌──────────┐ ┌─────────────┐ │
│ │ SQLite │ │ File bus │ │
│ │ WAL+FTS5│ │ heartbeat │ │
│ └──────────┘ └─────────────┘ │
│ ┌──────────────────────────┐ │
│ │ Monitor loop (30s poll) │ │
│ └──────────────────────────┘ │
└────────────┬────────────────────┘
│ spawn / bus read-write
┌────────┴────────┐
▼ ▼
Worker A Worker B
(git worktree) (git worktree)
operator/w-abc operator/w-defTwo-tier memory — tier-1 (volatile, auto-expire 30d) and tier-2 (durable, human-confirmed). Operator proposes classification; human confirms in dashboard (pending_classification=TRUE).
File bus — primary worker↔operator state channel. Workers write heartbeats to <bus-dir>/<worker-id>.json; monitor loop detects stale workers (no heartbeat >2min) and marks them failed.
Guardrails — JSON rules enforced before any tool call. Set via set_guardrail, inspected via get_guardrails.
Tool | Description |
| Create a task with goal, project path, priority, constraints |
| Launch a Claude Code worker (named or headless) on an isolated git worktree |
| Poll worker state, heartbeat age, current task |
| Mark task done, merge worktree branch, clean up |
| FTS5 full-text search across operator memory |
| Write/update a memory entry with tier + confidence |
| Add or update a guardrail rule |
| List active guardrail rules |
| Send an instruction to a running worker via file bus |
| Record a human judgment call for retrospective learning |
http://localhost:7373 — task board, worker status, guardrail editor, memory browser, judgment call log, SSE log stream.
claude-operator init # scaffold config + DB
claude-operator start # start MCP server + dashboard
claude-operator status # print active workers + tasks
claude-operator retro # open judgment call loggit clone https://github.com/MoyGulati/claude-operator
cd claude-operator
pnpm install
pnpm build
pnpm test # 39 tests
./verify.sh # full gate (build + test)Requirements: Node 24+, pnpm 10+.
MIT
This server cannot be deployed
Persistent, governed institutional memory for Claude Code — specs, decisions, learnings.
Persistent cross-session memory shared by Codex, Claude Code, ChatGPT, and other AI agents.
Adaptive plan/build/review cycles for AI coding assistants, persisted across sessions.
Build and supervise fleets of agents from Claude Code, Codex or Cursor. Connects over OAuth.
