Landlord

Parallel Claude agents with contracts, not prompts.

Landlord turns a single natural-language task into a plan of parallel Claude Agent SDK sessions — each bound by a contract (objective, checkpoints, JSON-schema outputs). Tenants that violate their contract get evicted and retried with fresh context. Every checkpoint output is validated against a JSON Schema and a structured LLM judge (via tool-use — no substring matching). Everything runs as an MCP server over stdio, so any MCP client (Claude Code, Cursor, Cline) can drive it.

Why

• No API credits required. Drives claude_agent_sdk.query(), which honors CLAUDE_CODE_OAUTH_TOKEN. Your Claude Pro/Max subscription runs decompose, tenants, and the judge.

• Contracts, not prompts. Structured output enforced by JSON Schema + LLM judge on every checkpoint. No "I hope the model said PASS."

• Tenants inherit your Claude Code config. Skills, CLAUDE.md, hooks, user MCP servers — all available inside every tenant via setting_sources=["user"] and skills="all".

• 5-tool MCP surface — start_orchestration, approve_plan, get_status, get_artifacts, cancel. That's the whole API.

• ~1,200 LOC runtime, 52 tests. Readable in an afternoon.

60-second install

pip install -e .
claude setup-token                       # one-time: get your OAuth token
setx CLAUDE_CODE_OAUTH_TOKEN "<paste>"   # Windows. Unix: export CLAUDE_CODE_OAUTH_TOKEN=...
claude mcp add -s user landlord landlord-mcp

Restart Claude Code. The five Landlord tools become discoverable; ask the model to orchestrate something.

Watch it work

Every tenant's SDK chatter streams to a tailable log:

tail -f ./landlord-output/<job_id>/job.json                   # orchestration state
tail -f ./landlord-output/<job_id>/<tenant_id>/session.log    # tenant model activity
ls  ./landlord-output/<job_id>/shared/                         # dependency artifacts

Auth details

Authenticate against your Claude Pro/Max subscription (no API credits needed):

claude setup-token

Set the resulting token in your environment:

# Windows (persistent)
setx CLAUDE_CODE_OAUTH_TOKEN "<token>"

# bash/zsh
export CLAUDE_CODE_OAUTH_TOKEN=<token>

(Advanced: if you'd rather pay per-token API usage, set ANTHROPIC_API_KEY instead — the underlying claude-agent-sdk accepts either.)

Related MCP server: AgentTasker MCP Server

Run the MCP server

landlord-mcp

This speaks MCP over stdio. Normally you don't run it directly — you point an MCP client at it. Easiest is the claude CLI:

claude mcp add -s user landlord <path-to-landlord-mcp-executable>

Or add this to ~/.claude.json (user scope) manually:

{
  "mcpServers": {
    "landlord": {
      "command": "landlord-mcp"
    }
  }
}

Restart Claude Code; the five Landlord tools will be discoverable to the model.

Tenant inheritance

Tenants spawned by the orchestrator run as Claude Agent SDK sessions with setting_sources=["user"] and skills="all". That means each tenant inherits:

• All user-level skills (invokable via the Skill tool)

• Your user CLAUDE.md (instructions/preferences)

• Your user-level MCP servers and hooks

• User memory

So a tenant can, for example, invoke /superpowers:writing-plans itself if your orchestrator decomposes "build feature X" into a tenant that needs to plan before coding. Per-contract skill allowlisting is a v2 feature; today it's all-or-nothing.

Tool surface

Configuration

Legacy CLI

The old litellm-based CLI is preserved as landlord:

pip install -e ".[legacy]"
landlord "your task"

See landlord/legacy/ for source.

Development

pip install -e ".[dev]"
pytest