Agent Lease

agent-lease-mcp

Keep two coding agents (Claude Code and Codex, or two sessions of either) from overwriting each other's work in the same repository.

agent-lease gives them leases on paths, presence and a shared message log — and the lease is enforced by a PreToolUse hook, not by good intentions. An agent that tries to edit a file another agent is working on is refused before the edit happens.

Česky / Czech version

Why it exists

9 September 2026. A coordination file, docs/AGENT-HANDOFF.md, existed and both agents knew about it. At 22:28 they still both touched deploy/make-work-package.sh: one was editing it while the other was running it. Bash reads a script as it goes, the text shifted under the running process, and instead of a 3.7 GB installer the build produced a useless 198 MB one.

The lesson the whole project rests on: coordination that can be ignored will be ignored. Not out of malice — an agent at work simply does not look at the notice board.

The same collision was later replayed on purpose: Codex ran the script, Claude Code tried to edit it mid-run, and the hook refused the edit. The file stayed intact.

How it differs

	locks	enforcement	TTL	presence	messages	audit
AgentRoom (paper)	✅	prompt only	?	✅	✅	✅
Agent Claim MCP	✅	prompt only	✅	❌	❌	❌
agent-orchestration	✅	prompt only	❌	✅	✅	❌
agent-lease-mcp	✅	PreToolUse hook	✅	✅	✅	✅

Advisory locks get broken; AgentRoom documents this itself. Claude Code and Codex both have PreToolUse hooks with the same semantics (exit 2 = block, reason on stderr), so a lease can be enforced deterministically at the tool-call level. One script serves both clients.

Rules it is built on

• A lease, not a lock. 30-minute TTL. A crashed agent cannot block the repo forever.
• All or nothing. A claim on several paths succeeds completely or not at all.
• A lease on a directory covers everything under it.
• A free file is claimed automatically on the first edit, so nobody has to remember to call claim.
• A running script is leased for the duration of the run — exactly the combination that broke the build.
• Fail-open by default. Broken coordination must not stop work; otherwise the first thing anyone does is switch the hook off. AGENT_LEASE_FAIL_CLOSED=1 flips this for unattended runs.

What runs on its own

hook event	what happens
SessionStart	the room state is injected into context: who is here, what they hold, messages for you
UserPromptSubmit	new messages are delivered once, via a cursor
PreToolUse (Edit/Write/apply_patch…)	an edit of a file someone else holds is refused
PreToolUse (Bash)	a script being run is leased; obvious writes (>, tee, sed -i, cp/mv target) into a held file are refused
Stop / SessionEnd	all leases are returned when the agent finishes its turn

Injections stay silent when there is nothing to say.

Tools

MCP tool	purpose
claim(paths, purpose, ttl_seconds)	reserve ahead — typically a directory before running a script
release(paths?)	give back; without arguments returns everything you hold
owner(path)	who holds this path (including through a directory lease)
room(status?)	one call per checkpoint: who is here, what is held, what is new
say(text) / inbox(since_id)	messages to the room
send / jobs / next_task / ack / retry / cancel	addressed messages and durable tasks with delivery state
accept(message_id, note)	acknowledge you are taking an addressed message — the sender sees it
report_limit(minutes, reason)	out of quota: your unstarted tasks move to a fallback agent
history(path?)	who touched what and how it ended — the answer to "why was I blocked"

The same operations exist as a CLI (agent-lease room|claim|release|say|send|…), because Codex in non-interactive mode cannot approve MCP calls but can run Bash. agent-lease wait --for <agent> blocks without spending model tokens and wakes a background session when a message arrives.

say/inbox is not a live chat: MCP is pull, the other side reads a message when it next looks. For a live view there is a local web room: agent-lease web --open (loopback only, see Security).

Quick start (two minutes)

uv tool install agent-lease-mcp
agent-lease init claude-code --write --rewake   # writes all hooks, keeps a .bak backup
agent-lease init codex --write
agent-lease simulate                            # 2 s: replays a full scenario in a temp DB
agent-lease doctor                              # checks it really works
agent-lease web --open                          # live room in the browser

init without --write only shows what it would write. Manual setup below.

Install

uv tool install agent-lease-mcp

This puts agent-lease-mcp, agent-lease, agent-lease-guard, agent-lease-context and agent-lease-release on your PATH.

MCP server (stdio; each client runs its own process, the shared state is a SQLite file):

# Claude Code
claude mcp add agent-lease -e AGENT_NAME=claude-code -- agent-lease-mcp

# Codex — ~/.codex/config.toml
[mcp_servers.agent-lease]
command = "agent-lease-mcp"
env = { AGENT_NAME = "codex" }

# Codex writes to ~/.agent-lease, outside the workspace — allow it:
[sandbox_workspace_write]
writable_roots = ["/home/YOU/.agent-lease"]

Hooks — without them this is just a polite notice board:

// ~/.claude/settings.json
{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Edit|Write|MultiEdit|NotebookEdit|Bash",
      "hooks": [{ "type": "command", "command": "AGENT_NAME=claude-code agent-lease-guard" }]
    }],
    "SessionStart":     [{ "hooks": [{ "type": "command", "command": "AGENT_NAME=claude-code agent-lease-context" }] }],
    "UserPromptSubmit": [{ "hooks": [{ "type": "command", "command": "AGENT_NAME=claude-code agent-lease-context" }] }],
    "Stop":             [{ "hooks": [{ "type": "command", "command": "AGENT_NAME=claude-code agent-lease-release" }] }]
  }
}

// ~/.codex/hooks.json — same events, AGENT_NAME=codex
{
  "hooks": {
    "PreToolUse": [{
      "matcher": "*",
      "hooks": [{ "type": "command", "command": "AGENT_NAME=codex agent-lease-guard" }]
    }]
  }
}

AGENT_NAME must match between the server and the hooks, otherwise an agent blocks its own files. An agent called unconfigured-* in room means exactly that.

variable	default	purpose
AGENT_NAME	unconfigured-<host>-<pid>	name in the room
AGENT_LEASE_DB	~/.agent-lease/room.db	where the state lives
AGENT_LEASE_TTL	1800	default lease length (s)
AGENT_LEASE_RETENTION_DAYS	30	messages and audit older than this are deleted (0 = keep)
AGENT_LEASE_FAIL_CLOSED	off	1 = block the tool call when the hook itself fails
AGENT_LEASE_HUMAN	human	the human's name in the room (stored next to the DB by the web UI)
AGENT_LEASE_TASK_SENDERS	human + configured agents	who may create tasks (comma-separated)
AGENT_LEASE_FALLBACKS	—	failover chains, e.g. codex=claude-code;claude-code=codex

Optional: Bash safety gate

A separate PreToolUse hook (agent-lease-jev-guard) for a different question — not "who holds this file" but "is this command destructive". Hard rules (rm -rf /, force push, git reset --hard, curl | sh, mkfs…) apply as soon as it is wired in. An optional external classifier (TypeSafe Jev) is off by default, starts in shadow mode, receives only redacted commands and is advisory — never the only safeguard. Details in ADR 0003.

Receipts, failover, bridge and broker

• Every addressed message has a visible state ⏳ → 👁 seen → ✔ taking → done. "Seen" is written by the hook automatically; "taking" must be acknowledged by the agent (accept) and is reminded on every prompt. Nothing gets lost silently. agent-lease overdue lists what nobody acknowledged.
• Limits and failover. When an agent runs out of quota (agent-lease limit or MCP report_limit), its unstarted tasks move to the next agent in AGENT_LEASE_FALLBACKS (e.g. codex=claude-code;claude-code=codex) after a grace period, and the sender is told.
• Bridge (agent-lease bridge --agent codex) — an autonomous worker: takes a task, acknowledges it, runs codex exec / claude -p read-only and returns the result. --provider openrouter --model <model> sends it to any model through OpenRouter as plain text without tools (read-only; good for review and a second opinion). Keys only from the environment.
• Broker (agent-lease broker) — single SQLite writer over a 0600 Unix socket; other tools use it automatically when it runs, so a sandboxed Codex does not need direct access to the database.
• Overview panel in the live room: each agent's active task, queue, unacknowledged items, limits and items to review. agent-lease notify --loop sends to Telegram only what waits on the human, each event at most once (TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID from the environment; --dry to test).
• systemd units in deploy/systemd/, Windows/WSL starter in deploy/windows/.

Details (in Czech): receipts, limits and failover, Jev.

Security

agent-lease is a coordination tool for agents running under one local user account. What that means in practice:

• Local trust model. Agent identity is the AGENT_NAME environment variable. Any process running as your user can claim to be any agent. The only reserved identity is the human one (michal in this build), which only the authenticated web UI can use. This protects against agents tripping over each other, not against a hostile process on your account.
• State is owner-only. room.db (and its WAL/SHM files) is 0600 in a 0700 directory; existing installs are tightened on start.
• Peer messages are untrusted data. In model context every message carries its kind, sender and recipient, is folded onto one line and is introduced with a note that it is information from a peer, not an instruction from the human. A task from another agent is a request, not an order.
• Tasks are bound to the caller. A peer sees only tasks it sent or received; lease tokens only for its own deliveries; retry/cancel need the sender or the recipient.
• Web UI binds to loopback only, exchanges a one-time URL-fragment token for an HttpOnly, SameSite=Strict session cookie, checks Origin and Host, and caps concurrent event streams.
• Bounds and retention. Message text, short fields, list limits and task lease length are bounded; old messages and audit entries expire, unfinished tasks never.

Known limits, stated plainly:

• Shell commands are checked best effort. Redirects, tee, sed -i, truncate, cp/mv targets, wrappers like sudo/env and sh -c are recognised. eval, variables in paths and arbitrary programs are not. This catches mistakes, not an agent that is trying to evade the hook.
• Check-then-use gap. The hook checks a path before the client opens it. A symlink swapped in between is not caught; a hook cannot bind the later file open to the checked inode.
• One machine. The state is a file on disk.

Found something? Please report it privately via GitHub Security Advisories.

Development

uv sync --extra dev
.venv/bin/python -m pytest -q
.venv/bin/ruff check .

Design decisions: ADR 0001 (hook enforcement), ADR 0002 (SQLite and stdio), ADR 0003 (separate Bash safety gate) — in Czech.

Built by Michal with Claude Code and Codex — the two agents it coordinates.

MIT licence.