SpecGate
GitHub

Spec governance for AI coding agents

Approve one version.
Verify the build.

Give coding agents approved context, then review delivery evidence against the same acceptance criteria.

Run locally View GitHub
WI-101 / audit-log permissions Reviewing…

Acceptance criteria / evidence

  • Admins can view permission-change events
  • Member roles cannot read restricted payloads
  • Tests cover allow and deny paths
tests types lint
delivery review / per-criterion evidence

Your tools stay in the loop.

Install through the CLI, or use the Codex and Claude Code plugin managers.

CLI and IDE handoff

Code where you already work

Your agent reads the approved Context Pack and returns delivery evidence through the CLI.

Coding agents
  • Cursor
  • Claude Code
  • OpenAI Codex

Full mode

Repositories and optional work tracking

GitHub and GitLab confirm merged PRs and MRs at the submitted commit. Linear can receive approved work without replacing direct IDE pickup.

Repositories
  • GitHub
  • GitLab
Work tracking
  • Linear

The governed loop in three moves.

SpecGate does not replace your spec tool, tracker, or IDE. It makes the handoff explicit, then checks the finished work against the approved contract.

  1. Publish

    Bring any spec format

    OpenSpec, Spec Kit, Superpowers, a quick change note, or your own document set. You declare the document roles and SpecGate resolves the automatic policy.

  2. Hand off

    Approve one Context Pack

    The coding agent gets the approved version, acceptance criteria, non-goals, risks, tasks, and applicable Skills through the CLI.

  3. Review

    Review delivery evidence

    SpecGate compares completion claims, tests, docs, and changed files against every acceptance criterion before you close the work.

15-second product tour

When "done" is not delivered.

A coding agent reports completion. One acceptance criterion has no trusted evidence. SpecGate keeps the work open until the gap is fixed, reviewed, and accepted.

01Agent says done 02Evidence gap found 03Human accepts the fix

See it run, command by command.

The CLI drives every SpecGate workflow. Cursor, Claude Code, and Codex use the same Context Pack and report evidence the same way.

specgate / approved artifact 
$ specgate artifact publish --file artifact.json --preview
  Artifact publish preview:
  docs/spec.md	 spec	 4821 bytes
  No publication performed: Human confirmation required before publishing.
$ specgate artifact publish --file artifact.json
  Published 8a42…
$ specgate gates check <artifact-id>
  Spec quality: pass
$ specgate --yes artifact approve <artifact-id>
  Approved 8a42… (v1)

Approve the contract. Publish immutable documents, run readiness, then record the human decision.

Keep your stack. Add the governed handoff.

SpecGate connects planning and implementation. Author where you already work, approve one version, then review the delivery against it.

SPEC TOOLS

Shape the work

OpenSpec, Spec Kit, Superpowers, Markdown, and custom docs stay useful as authoring systems.

TRACKER + IDE

Run execution

Issues, sprints, code changes, PRs, and day-to-day collaboration stay in your existing stack.

SPECGATE

Own the governed handoff

Versioned approval, readiness gates, Context Packs, delivery review, and evidence trust live here.

  • Approved spec versions
  • Readiness gates
  • Context Pack handoff
  • Delivery review verdicts

What teams ask first.

How is this different from spec tools?

Spec tools help shape work. SpecGate versions the result, runs readiness gates, hands one approved Context Pack to the coding agent, and reviews delivery evidence against it.

What spec formats do you accept?

Bring an OpenSpec, Spec Kit, or Superpowers project, Markdown, or your own document set. You explicitly map each document to a role and SpecGate resolves the automatic policy.

Do we have to leave our coding agent?

No. Cursor, Claude Code, Codex, or a human can use the CLI. The agent reads the Context Pack before coding and reports evidence after implementation.

Do I need an LLM API key?

No. The deterministic core works without a model. Configure one only when you want independent model-backed readiness or delivery judgment.

What does the delivery review check?

Each acceptance criterion is checked against reported evidence and automated checks such as tests, types, and lint. Failed criteria stay visible for the next handoff.

How should we roll this out?

Start with the Local CLI: one binary, no Docker, no server, and your governance data stays on your machine. Add the appliance when you want the browser UI, governance chat, or shared workspaces. It is built for a trusted machine or private network, so put authentication and TLS in front of it before exposing it beyond that.

Close the loop on agent delivery.

The local-first default uses no Docker or server. Hand off approved context and review the evidence that comes back.

Run locally View GitHub