What runs today
Every capability the site and setup mention, in one of four classes, with the evidence for each.
Where it is supported
The beta runs on GitHub.com, hosted by MergeReady, through one GitHub App installed once. Every repository starts at Observe: MergeReady reads, triages and keeps one pack comment and one check per PR, and writes nothing to code. Propose and Act are raised per path by an owner or admin, in the policy MergeReady keeps.
Other hosts, the command line, the MCP server and the self-hosted control plane are built and tested, and each is offered once it passes its live check. Approve and Merge stay off, behind their own checks.
The four classes
- Verified, on by default. Tested in the repository, and checked live on GitHub.com in a recorded run. It works from the install.
- Verified, you turn it on. The same evidence. It is off until your policy or an owner turns it on.
- Built, not yet verified. The code and its tests are in, and a live check or a bar it must clear has not passed yet. It is not offered as available today, however complete it is.
- Planned. On the roadmap, not built yet.
Refusals and limits (protected paths, the caps, no keys in the sandbox, never approving its own push) are not capabilities: they hold at every level and are tested in the repository.
Every capability, by class
Evidence names records in MergeReady's source repository, which is not public yet; the release record below publishes what they establish.
| Capability | Where | Evidence | Still to pass |
|---|---|---|---|
| Verified, on by default | |||
| One GitHub App, installed once, with every repository at Observe | GitHub.com | One App made from the manifest and installed on a test organization; single-repository tokens minted, scoped and revoked: docs/plan/live-github-2026-11-05.md. docs/adr/0127-one-github-app.md, packages/hosts/test/manifests.test.ts | |
| The evidence pack: one comment per PR, edited in place | GitHub.com, at Observe | One pack per PR, edited in place, through the deployment: docs/plan/live-github-2026-11-06.md, docs/plan/live-github-2026-11-05.md | |
| The Merge readiness check, which never fails a PR on the agent's opinion | GitHub.com, at Observe | docs/plan/live-github-2026-11-06.md, docs/plan/live-github-2026-11-05.md, packages/hosts/test/github-check-status.test.ts | |
| Triage: the PR's size and risk, in the pack's first revision | GitHub.com, at Observe | The production triage agent on a live model, injection cases included: docs/plan/live-models-2026-11-03.md; runs through triage on the deployment: docs/plan/live-github-2026-11-05.md | |
| One named reviewer asked, and asked again only when there is something new | GitHub.com, at Observe | The CODEOWNER requested through the deployment, and the second request it found fixed: docs/plan/live-github-2026-11-06.md | |
| Verified, you turn it on | |||
| Split proposals: a large PR proposed as stacked PRs, each green | GitHub.com, from Propose | The split question, stacked PRs each passing CI, and a second accept opening nothing: docs/plan/live-github-2026-11-06.md | |
| Proposed tests, as a stacked PR for a person to apply | GitHub.com, from Propose | A tests PR at ready, through the cross-family verifier: docs/plan/live-github-2026-11-06.md | |
| Private registries, installed from through the proxy with no credential in the sandbox | GitHub.com, set up by an owner | An install from a private registry in a hardened session: docs/plan/live-github-2026-11-06.md | |
| Built, not yet verified | |||
| Any author: agents recognised by bot login, co-author trailer or branch prefix, a person when unsure | GitHub.com, at Observe | packages/policy/src/authors.ts and its tests; no recorded live run shows an agent's PR recognised | A live PR opened by an agent, recognised as one, on the release candidate (#622) |
| Pre-flight: the affected tests in a sandbox before CI, on by default at every level | GitHub.com, every level | packages/policy/src/schema.ts (ci.preflight, on); agreement with CI on a recorded corpus: docs/plan/live-github-2026-11-06.md | Pre-flight in the hosted sandboxes on the release candidate, after their hardening (#609, #622) |
| The linked issue's criteria, each marked covered or not | GitHub.com, at Observe | packages/agent/src/intent and its tests | The intent oracle's calibration on real PRs (#318) |
| Duplicate detection, and closing a duplicate agent PR after a notice when the policy turns it on | GitHub.com, at Observe | Duplicates right on 5 of 5 samples: docs/plan/live-models-2026-11-03.md | Precision on the labelled sets (#317) |
| The PRs a change collides with | GitHub.com, at Observe | Tests in the repository | No recorded live run yet |
| Nudges to whoever a PR is waiting on | GitHub.com, at Observe | Tests in the repository | No recorded live run yet |
| Closing an agent's PR left 14 days waiting on an answer or a review | GitHub.com, at Observe | Tests in the repository | No recorded live run yet |
| Act: verified fixes for formatting, lint and test failures, committed by your host | GitHub.com, from Act | A red PR fixed to ready through the deployment four times: docs/plan/live-github-2026-11-05.md; 13 of 14 fixed with 0 bad pushes on the hand-built set: docs/plan/live-models-2026-11-03.md | Gate A, its first bar: at least 30% of 200 replayed real CI failures fixed with zero bad pushes, not yet measured (docs/plan/gate-a.md, #319) |
| Propose's fixes: suggested patches and stacked fix PRs a person applies | GitHub.com, from Propose | Tests in the repository | The fixes' bar, Gate A (#319); no recorded live run of a stacked fix PR |
| Cancelling the CI runs its own push replaced | GitHub.com, from Act | A cancel sent through the deployed relay: docs/plan/live-github-2026-11-06.md | With Act's fixes, Gate A (#319) |
| Review comments: the change a clear comment asks for, the reply, the resolved thread; rebases | GitHub.com, from Act | Every phase-2 flow replays: docs/plan/gate-b.md | Gate B: the comment and conflict bars and two weeks with design partners (#310, #311, #336) |
| Reconciling other review bots: duplicate findings merged, those below your floor listed as ignored | GitHub.com, from Act | docs/plan/gate-b.md (the trusted bot's comment flow) | The comment bar (#311) |
| Flaky tests quarantined in an issue and rerun, never edited | GitHub.com, from Propose | Flake precision met on recorded projects: docs/plan/live-github-2026-11-06.md | No recorded live quarantine on a PR from start to end |
| Questions in chat: one person asked with a button per answer, the channel after two days | GitHub.com with Slack | Tests in the repository | Slack's consent, uninstall and message withdrawal, live (#333) |
The command line, mergeready, on your machine with your own model keys | GitHub.com | Installed and run as a user: docs/plan/cli-local-smoke-2026-11-03.md; run --apply on a test PR: docs/plan/live-github-2026-11-05.md | Its publication, so a visitor can install it (#334) |
| The MCP server coding agents call before and after they open a PR | GitHub.com | A real Claude Code session and a scripted client against the deployment: docs/plan/live-github-2026-11-06.md | Its publication with the command line (#334) |
| Running it in your own CI: a read-only job and an apply job | GitHub Actions | The apply job pushed a fix as the customer's own App on a same-repository PR: docs/plan/live-github-2026-11-06.md | The command line's publication (#334) |
| The web app: setup, the policy editor, the worst-first queue and pause | GitHub.com | apps/web/test/setup.test.tsx, apps/web/test/e2e-setup.test.ts, apps/web/test/e2e-queue.test.ts; pause and the queue through the MCP server: docs/plan/live-github-2026-11-06.md | One continuous customer journey on the release candidate (#622) |
| Your own model keys, stored once and used for your organization's runs | GitHub.com | Tests in the repository | The customer journey on the release candidate (#622) |
| Audit export: the log of every write and refusal as JSON Lines | GitHub.com | Tests in the repository | The customer journey on the release candidate (#622) |
| Approve and Merge: approving docs-only PRs after a veto window, and entering the merge queue; off by default | GitHub.com, with your org's allowance | A docs-only PR approved after its veto window and merged by a real merge queue: docs/plan/live-github-2026-11-06.md | Their own checks, outside this launch (#624, #625) |
| GitLab.com and self-managed GitLab, through a group or project access token | GitLab | packages/hosts/test/gitlab.test.ts, against fakes | Contract runs, CI jobs and merge-SHA binding on a live GitLab group (#329) |
| Bitbucket Cloud, through a workspace or repository access token | Bitbucket Cloud | packages/hosts/test/bitbucket.test.ts, against fakes | A contract run on a live workspace (#330) |
| GitHub Enterprise Server at Observe, on the self-hosted control plane | GitHub Enterprise Server | packages/hosts/test/ghes.test.ts | An install on a licensed GitHub Enterprise Server (#337) |
| The self-hosted control plane in your VPC, from the Helm chart | Your Kubernetes cluster | Installs, migrates, answers health and upgrades on a local kind cluster: docs/plan/live-github-2026-11-06.md; deploy/helm/mergeready | The GitHub Enterprise Server half (#337) |
| Planned | |||
| SSO and SCIM for enterprise sign-in | — | Not built | |
| Hosts beyond these: Bitbucket Data Center, Azure DevOps and others, in the order the waitlist asks for them | — | Not built | |
The release record
Recorded, with no release behind it yet. The capability ledger kept with the source (docs/plan/release-capability-ledger.md) is bound to a source commit that is not tagged, released or deployed. Its images, configuration and environment read "not deployed" until a release fills them.
No live check has run on that source, so no row above is verified on it. Verified still means the recorded checks named above, on the earlier builds they name.
The launch decision reads blocked, for an invited beta and for open sign-up alike, until the live checks and owner actions the ledger lists in order are done. The owner has not reviewed it yet.