Integrates with GitHub as the shipping provider, enabling APE to push run branches, open or reuse pull requests, wait for required checks, and squash-merge with optional auto-merge.
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., "@ape Start a phase run to add optimistic locking to invoices"
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.
APE runs coding tasks through planning, tests, implementation, review, and shipping. It saves progress outside the chat, so a task can continue in another session.
APE coordinates the host's native agents. Its runtime decides the next stage and checks the evidence before accepting a worker's result. This helps catch unsupported claims; it does not guarantee that a test, review, or code change is correct.
Version 2.25.8 is approved for release with a documented owner exception. The tested candidate passed all 18 CI jobs and completed the mechanical, fast, full, and protected-land Codex workflows. The full workflow required recovery after two host stream interruptions and an expired, unspawned test ticket; this is not an uninterrupted schema-v5 first-pass certificate. Runtime and plugin package bytes are preserved from the tested candidate. See the release notes and evidence and limits.
Codex CLI is the primary host. The Claude Code package is included, but Claude live operation is unverified. Codex IDE integrations and ChatGPT web, mobile, and cloud runtimes are not supported. Codex is the sole required live release-certification host; a Claude subscription or API account is not required to use APE with Codex or publish its release.
Installing APE does not start a run. You choose when it can work and ship.
Related MCP server: claude-swarm
Node.js 22.12.0 or newer is required.
Use the pinned Codex CLI 0.153.4. Install the plugin and enable the native V2 worker tools in the Codex home you use for APE:
codex plugin marketplace add AAWWCC/ape
codex plugin add ape@ape
codex features enable multi_agent_v2The feature command updates that Codex home's config.toml. Restart Codex, then
use codex features list in the same launch environment to verify that
multi_agent_v2 is true. The older multi_agent setting alone does not ensure
the worker interface APE requires. See host compatibility
for the equivalent configuration and launch option.
Run inside Claude Code:
/plugin marketplace add AAWWCC/ape
/plugin install ape@ape
/reload-pluginsBoth packages run locally through Node.js. APE stores project state under
.ape/runtime/; it does not send that state to an APE-operated service.
The coding host and configured tools still use their own services and permissions.
compatibility.json defines the supported host versions
and platforms. See host compatibility before changing a
CLI version. CI checks packages on Linux, macOS, and Windows.
Start by setting up the project's test commands. Configuration changes are proposed for your approval:
/ape:config init
/ape:config doctorThen start a task, check its progress, or continue it later:
/ape:run Add optimistic locking to invoice updates
/ape:status
/ape:resumeUse APE for work that needs tests, review, or continuity across sessions. For a tiny edit or a quick question, an ordinary coding session may be simpler.
Skill | What it does |
| Start a build, investigation, or shipping run. |
| Show progress, pending work, and blocks. |
| Continue an interrupted run. |
| Inspect past runs and their results. |
| Set up commands, models, and shipping. |
| Request an audited abort, reset, or dispatch expiration. |
| Manage an optional project roadmap. |
Only read-only status may be selected automatically. Invoke the other skills
explicitly. See the skill reference for details.
Mode | Work |
| Plan, test, implement, review, and ship a change. |
| Investigate a problem without editing files. |
| Research a question without editing files. |
| Review and ship existing work without editing it. |
A phase run uses one of three lanes:
Mechanical: documentation, generated files, and other non-behavioral changes.
Fast: small behavioral changes without high-risk triggers.
Full: larger or sensitive changes, such as authentication, migrations, or public APIs.
Choose auto to let APE classify the lane. A run can move to a stricter lane,
but not a lighter one.
Behavioral changes normally start with failing tests, written by a separate
test worker. Use green-maintenance for tests that should already pass, such as
a regression net or deflake. Non-behavioral work does not need fabricated failing
tests. See pipeline rules for scope limits and stage order.
Before dispatching workers, APE checks the proposed scope, repository state, commands, required capabilities, and later pipeline stages. Missing paths need approval; they are not silently added.
Workers must return validated results called receipts. APE also checks file scope, tree identity, tests, reviews, and the configured gates before shipping.
GitHub is the only shipping provider. Each project needs its own explicit
repository target. By default, green work waits for an audited ship action.
With shipping.auto_merge: true, a run also needs explicit per-run shipping
approval before APE may push, open a PR, and merge.
APE waits for required checks and proof of the remote merge. A failed local cleanup does not undo a proven remote success. A blocked run does not authorize automatic reset, abort, or restoration.
See configuration and pipeline rules for the exact checks and recovery options.
npm ci
npm run public:hooks
npm run typecheck
npm run test:agent
npm run bundle
npm run package:plugins
npm run package:check
npm run public:checknpm test uses six workers. npm run test:agent -- <paths...> uses three and is
the preferred profile when other agents may also be testing.
public:check scans both packaged plugins and the actual Git worktree inventory,
including new nonignored source files. Tracked docs and tests remain checked even
when ignore rules change. Negative tests construct their explicitly synthetic
protected-looking inputs from fixture components; no source directory is exempt.
Never place real private values in source, including split or encoded forms.
Keep runtime changes, tests, and generated packages easy to review separately. The Git hooks check commit identities; verify the remote before pushing. A rebuild does not update an installed plugin. Reinstallation, pushes, and publication are separate actions.
The package and release builders accept --output-root for a new or empty directory,
or one containing only recognized APE output. They preserve unfamiliar files and refuse
source directories. A package upgrade that removes an old generated filename may need
a fresh output directory; inspect the retained files before removing them yourself.
For the full checklists, see Contributing, release checks, prompt evaluations, and performance baselines.
Security policy for private vulnerability reports
APE is licensed under MIT. See third-party notices for material with separate terms.
This server cannot be deployed
Adaptive plan/build/review cycles for AI coding assistants, persisted across sessions.
Cross-agent artifact workspace with provenance across Claude Code, Codex, Cursor, LangGraph.
Coordinate coding agents through MCP using existing AI plans, saved work, and independent checks.
Source-checked CLI guides and model-aware planning for Claude Code, Codex, and Grok Build.
