agentctl

Let your coding agent delegate to other coding agents.

Ask Claude Code to get a review from Codex, or have Codex hand a bug to Grok in Cursor. agentctl launches the other agent's CLI, tracks the run, and brings the answer back, even if the agent that asked has moved on or lost its context.

You, in Claude Code:

  > Have Codex review this branch, then send its findings back and ask it to fix them.

What Claude runs:

  $ agentctl delegate --request-file review.json --prompt-file review.md --wait --content
  exec-snack-need-scrap-group-correct-marriage state=completed harness=codex
  <Codex's review>

  $ agentctl continue exec-snack-need-scrap-group-correct-marriage \
      --request-key fix-01 --prompt-file fix.md --wait --content
  <the same Codex session, now applying its fixes>

You talk to your agent in plain language. agentctl ships a skill that teaches the agent these commands, so you never write the JSON yourself.

Pre-1.0: commands, JSON output, and config formats may still change.

Why use it

• Ask for a model by name. "codex", "grok", or "claude" resolves to the right CLI, model, and flags from your config. Your agent doesn't need to know any other CLI's flags.
• Set it up once, use it on every machine. Keep your agent and model choices in one Git repository. Each machine pulls the same config, and the installer adds the agentctl skill to every agent CLI it finds.
• Work can outlive its parent. delegate --detach returns a tracked execution ID while an independent native owner runs the task. Collect once with await <id> --content. See the lifecycle contract for recovery and authority limits.
• Answers don't get lost. Every run gets an ID and a stored result. If the asking agent times out, crashes, or starts a new session, agentctl inbox shows what finished and agentctl result returns the answer. Retrying with the same request key returns the original run instead of starting a second one.

agentctl doesn't replace your agent CLIs, proxy model APIs, or store your credentials or transcripts. Each CLI keeps its own login, permissions, and sessions.

agentctl children discovers retained attributed work for the current caller. Its completeness claim covers stable membership of that retained index; mutations between pages fail closed. Lost remote receipts still require identical-key recovery, and existing parent-graph retention remains in force. See child work and its coverage boundaries.

Supported agents

Agent	Executable
Claude Code	claude
Codex	codex
Cursor	cursor-agent
Devin	devin
Hermes	hermes
ZCode	zcode
OMP	omp
OpenClaw	openclaw
OpenCode	opencode
Any CLI with structured output	configured

Follow-up turns and steering depend on what each CLI supports; agentctl capabilities <agent> reports the current answer. agentctl can also dispatch work to Multica agents on other machines.

Install

Download the archive for your platform and SHA256SUMS from the latest release, verify it, and run the bundled installer:

grep " agentctl_<version>_darwin_arm64.tar.gz\$" SHA256SUMS | shasum -a 256 -c -
tar -xzf agentctl_<version>_darwin_arm64.tar.gz
cd agentctl_<version>_darwin_arm64
scripts/install.sh --binary "$PWD/agentctl"
agentctl doctor

Releases are built for macOS and Linux on arm64 and amd64. The binary goes in ~/.local/bin, and the installer adds the agentctl skill to each agent CLI it detects. agentctl updates itself once a day; agentctl update policy notify or off changes that.

To build from source instead, you need Go:

git clone https://github.com/Git-on-my-level/agentctl.git
cd agentctl
make ci
go build -o build/agentctl ./cmd/agentctl

Configure your agents

List the agents and models you want to delegate to in ~/.config/agentctl/config.json. Only listed entries can be launched by name, and default breaks ties when a name matches more than one entry:

{
  "schema_version": 1,
  "default_profile": "local",
  "profiles": {
    "local": {
      "agent_preferences": {
        "mode": "advisory",
        "preferred": [
          {"agent": "codex", "family": "gpt", "model": "<codex model>", "speed": "regular", "default": true},
          {"agent": "cursor", "family": "grok", "model": "<cursor model>", "speed": "regular"},
          {"agent": "claude-code", "family": "claude", "model": "<claude model>"}
        ]
      }
    }
  }
}

agentctl config doctor

To share this across machines, commit the same JSON as config-bundle.json in a Git repository and point each machine at it:

agentctl config source init --remote git@github.com:you/agentctl-config.git
agentctl config source update   # pull later changes

agentctl adds flags that skip an agent's approvals or sandbox only when your config grants it. See Structured delegation for permissions, workspace trust, and matching rules, and Configuration for everything else.

Learn more

agentctl help lists every command; agentctl help <command> explains one.

• Structured delegation
• Follow-up turns and steering
• Running several agents at once
• Adapters
• Architecture
• State, security, and privacy
• Roadmap

To contribute, see Contributing. Report vulnerabilities as described in Security.

License

Apache License 2.0. Third-party notices are in NOTICE.

Caller attribution and bounded read-only child discovery: child work.