Invoker

Persisted multi-agent workflow orchestration

A DAG of tasks in isolated git worktrees, composed through merge gates and review — desktop, CLI, and Slack on one control plane.

Download · Website · Join Slack

Watch the Invoker demo video

Features

Monitor execution

See parallel runs, dependencies, PRs, and replay paths in one stacked workflow graph — persistence is the source of truth, not process memory.

Drive with AI

Plan and run Codex, Claude, and other agents as first-class DAG nodes with explicit lineage and session audit trails.

Intervene and approve

Human gates are first-class states — approve, retry, or redirect without leaving the control plane.

Review work

Branches, merges, conflicts, and pull requests are part of the execution model — review stacked changes in context.

Control cloud and remote agents

Spread work across SSH targets and remote machines you already manage — same actions from desktop, CLI, or Slack.

Workers

Invoker background work is owned by the built-in worker registry. autofix is the default failed-task recovery worker; it reconciles persisted state and submits normal fix-with-agent intents instead of running recovery from task-state producers.

Run workers on the Invoker owner host. Operators can start and stop them from the desktop Workers tab, inspect them with invoker-ui --headless worker status --output text|json|jsonl, or trigger one explicit scan with invoker-ui --headless worker <kind>. Each kind takes its own single-instance lock, so a second scan of the same worker is refused without blocking other worker kinds.

PR maintenance uses the same owner-host worker path. Enable prMaintenance to launch pr-admin-bypass-land, keep pr-orphan-repair available from the same built-in registry, and do not install separate cron jobs or external worker launchers for the supported setup.

Install

Requires Node.js 26.x. One command installs CLI + UI, runs doctor, and turns on pr-status / autofix / auto-approve (skips Slack, remote machines, and harness skill/MCP wiring):

npx @neko-catpital-labs/invoker-cli@latest install

Sample output: docs/install-transcript.txt (invoker-cli install --demo).

Default Invoker owner is local. In chat, name a host or IP to use a remote Invoker (the agent probes SSH and retargets MCP). Slack and remote machines stay optional (invoker-cli setup slack / setup machines).

Manual equivalent:

npm install -g @neko-catpital-labs/invoker-ui
npm install -g @neko-catpital-labs/invoker-cli
invoker-cli install

If you need Node installed first, optional wrapper: curl -fsSL https://raw.githubusercontent.com/Neko-Catpital-Labs/Invoker/master/scripts/bootstrap.sh | bash (ensures Node 26, then runs the same npx @neko-catpital-labs/invoker-cli@latest install). Desktop packages only: curl -fsSL https://raw.githubusercontent.com/Neko-Catpital-Labs/Invoker/master/scripts/install.sh | bash. Full install, config, and source checkout steps: Getting started.

invoker-cli install enables default workers (pr-status, autofix, auto-approve). It does not install skills under ~/.invoker or modify Cursor / Claude / Codex / OMP configs by default. Interactive invoker-cli setup asks whether to wire those harnesses (default No); pass --register-harnesses to opt in non-interactively (--yes alone does not register). Setup still walks Slack and machines when you want them.

After harness registration, ask in Codex, Claude, Cursor, or OMP to plan and run durable work through Invoker. The invoker-chat-submit skill and MCP tools prepare a review, wait for one approval, submit, and watch status without a slash command.

Explicit fallback remains available:

/invoker-plan-to-invoker "help me plan <change>"

Either path writes plans/invoker-handoff.md, converts it to plans/invoker-handoff.yaml, validates, and reviews via MCP (invoker_prepare_plan_review → one approval → invoker_submit_plan with reviewToken), then reports status with invoker_get_workflow / invoker_list_tasks / bounded invoker_wait_for_workflow. The explicit fallback command can also submit with invoker-cli run --live instead of the Invoker MCP tool.

Docs

• Getting started — prerequisites, install, config, quick start, troubleshooting
• Architecture — package layering, mutation boundaries, error contracts
• invoker-ops skill — operate existing workflows: retry, restart, cancel, or inspect blocked tasks
• Slack native workflows — lobby mentions, harness presets, per-workflow channels
• First agent workflow tutorial — guided first run
• Changelog
• Contributing
• License

More: local macOS release build, remote SSH targets, Docker executor, web surface, UI/backend drift tracing, product story.

If you need to turn a product or implementation plan into an Invoker workflow, run invoker-cli setup (or System Setup in the desktop app) to install helpers. In normal chat, ask to plan/submit durable work through Invoker; the installed invoker-chat-submit skill and MCP tools handle review, one approval, submit, and status. /invoker-plan-to-invoker "help me plan <change>" remains the explicit slash-command fallback.

License

Functional Source License, Version 1.1, ALv2 Future License (SPDX: FSL-1.1-ALv2). Permitted use, competing use, and the future Apache License 2.0 grant are defined in the license file.

Invoker also includes the Neko Catpital Ventures, LLC Addendum in LICENSE. In plain terms, that addendum says:

• if you modify or redistribute the Software for commercial use, those modifications or redistributions must remain open source under the FSL and the NCV Addendum, and cannot be relicensed more restrictively
• you may build and exploit software or developments using Invoker, so long as Invoker itself is not incorporated into that software or those developments
• except for evaluation or testing, you may not use the Software to replace employees or reduce headcount for substantially similar roles for six months after first production use

The LICENSE file is the controlling text, including the full NCV Addendum.