Crit Integrations

September 8, 2026 · View on GitHub

Drop-in configuration files that teach your AI coding tool to use Crit for reviewing plans and code changes.

Quick install

crit install <tool>     # Install for a specific tool in the current project
crit install all        # Install for all supported tools

Safe to re-run. Existing files are skipped (use --force to overwrite).

Global install: run cd ~ && crit install <tool> to install to your home directory. The integration is then available across all projects without per-project setup. Each tool reads from a different global path; crit install routes the files to the right place automatically.

Tool	Install command	Project destination	Global destination
Claude Code	`crit install claude-code`	`.claude/skills/crit/SKILL.md` + `crit-cli` + `crit-story`	`~/.claude/skills/crit{,-cli,-story}/SKILL.md`
Cursor	`crit install cursor`	`.cursor/skills/crit/SKILL.md` + `crit-cli` + `crit-story`	(project only — Cursor has no stable user-level config dir)
GitHub Copilot	`crit install github-copilot`	`.github/skills/crit{,-cli,-story}/SKILL.md`	`~/.agents/skills/crit{,-cli,-story}/SKILL.md`
OpenCode	`crit install opencode`	`.opencode/commands/crit.md` + `crit-story.md` + skills + plugin	`~/.config/opencode/commands/` + `~/.agents/skills/` + plugins
Codex	`crit install codex`	`.agents/skills/crit{,-cli,-story}/SKILL.md`	`~/.agents/skills/crit{,-cli,-story}/SKILL.md`
Codex plugin	`crit install codex-plugin`	loose skills + marketplace + `plugins/crit/` (incl. crit-story)	`~/.agents/` + `~/.codex/plugins/crit/`
Pi	`crit install pi`	`.pi/skills/crit{,-cli,-story}/SKILL.md`	`~/.pi/agent/skills/crit{,-cli,-story}/SKILL.md`
Qwen Code	`crit install qwen`	`.qwen/skills/crit{,-cli,-story}/SKILL.md`	`~/.qwen/skills/crit{,-cli,-story}/SKILL.md`
Hermes	`crit install hermes`	`.hermes/skills/crit{,-cli,-story}/SKILL.md` (add `.hermes/skills` to `external_dirs`)	`~/.hermes/skills/crit{,-cli,-story}/SKILL.md`
Windsurf	`crit install windsurf`	`.windsurf/workflows/crit.md` + `crit-story.md` + skills	`~/.codeium/windsurf/global_workflows/` + skills
Cline	`crit install cline`	`.clinerules/workflows/crit.md` + `crit-story.md` + skills	`~/.cline/data/workflows/` + `~/.cline/skills/`
Aider	`crit install aider`	`.crit/aider-conventions.md` + adds entry under `read:` in `.aider.conf.yml`	`~/.crit-conventions.md` + adds entry under `read:` in `~/.aider.conf.yml`
Gemini CLI	`crit install gemini`	`crit-cli` + `crit-story` skills + `/crit` + `/crit-story` commands + policy	same under `~/.gemini/`
Grok	`crit install grok`	`.grok/skills/crit{,-cli,-story}/SKILL.md`	`~/.grok/skills/crit{,-cli,-story}/SKILL.md`
Amp	`crit install ampcode`	`.agents/skills/crit{,-cli,-story}/SKILL.md`	`~/.config/agents/skills/crit{,-cli,-story}/SKILL.md`

Plugin marketplace (Claude Code)

For the full experience, install via the plugin marketplace. This gives you:

A /crit slash command for the review loop
A model-discoverable crit-cli skill for review files, crit comment, crit pull/push, etc.
A wording-gated /crit-story skill for chaptered diff overviews that then continues the /crit review loop

claude plugin marketplace add tomasz-tomczyk/crit
claude plugin install crit@crit

The marketplace manifest lives at the repo root (.claude-plugin/marketplace.json) and points to the plugin files in integrations/claude-code/.

`crit install` vs plugin marketplace

	`crit install`	Plugin marketplace
Scope	Per-project (committed to repo)	Global (user-wide)
What's installed	`/crit` skill only	`/crit` skill + `crit-cli` skill
Good for	Teams — everyone gets the integration	Individual users — works across all projects
Setup	Run once per project	Install once, works everywhere

Both approaches give you the user-invoked /crit review cycle. The plugin marketplace additionally installs the crit-cli skill, which can auto-teach the agent about crit comment, review file format, crit pull/push, and resolution workflow without starting the interactive browser loop.

Claude Code plan approval mode

The Claude Code plugin intercepts ExitPlanMode with a narrowly matched PermissionRequest hook. By default, approving in Crit allows the plan exit and leaves Claude Code to restore its existing permission mode. To choose the mode deterministically, set plan_approve_mode in your global Crit config:

{
  "plan_approve_mode": "acceptEdits"
}

Disable automatic plan review per shell or globally with export CRIT_PLAN_REVIEW=off. This only disables the plan-exit hook; manually running crit plan still opens a review.

Supported values are default, manual, acceptEdits, plan, auto, dontAsk, and bypassPermissions. The manual alias requires Claude Code 2.1.200 or newer. On approval, Crit returns Claude Code's documented decision.updatedPermissions entry:

{
  "type": "setMode",
  "mode": "acceptEdits",
  "destination": "session"
}

destination: "session" keeps the change in memory for the current Claude Code session only. The setting is global-only (~/.crit.config.json); a repository's .crit.config.json cannot change your permission policy. Unset preserves the default hook behavior, and invalid values are ignored with a warning.

The hook approval and mode switch do not override matching deny or ask rules. Claude Code can also disable auto through permissions.disableAutoMode. bypassPermissions is intentionally dangerous and should only be used in an isolated environment. Claude Code applies it only when the session started with bypass mode available (for example --allow-dangerously-skip-permissions or --dangerously-skip-permissions) and managed settings have not disabled it; otherwise Claude Code treats the update as a no-op.

crit install opencode also writes a small TypeScript plugin (crit.ts) and registers it in opencode.jsonc. The plugin shells out to crit config on each chat turn and appends sharing instructions to the system prompt only when share_url is set. With share_url: "" the sharing block is omitted entirely — useful in environments with strict information-sharing policies, and saves tokens otherwise. opencode auto-loads .ts files dropped into the plugin directory, so the registration entry is informational.

Codex plugin

For the full Codex experience, install the plugin. This gives you:

A $crit skill for the review loop (plus loose copies under .agents/skills/ so bare $crit works even outside the plugin)
A crit-cli skill that auto-activates when working with review files, crit comment, crit pull/push, etc.
A proposed-plan review hook — intercepts Codex's Stop hook when the agent proposes a plan in Plan mode, writes it to disk, and opens Crit for inline review before the turn ends

cd ~ && crit install codex-plugin    # global (recommended)
crit install codex-plugin            # per-project (commit plugins/crit/ for the whole team)

crit install codex-plugin registers the plugin in a local Codex marketplace (.agents/plugins/marketplace.json or ~/.agents/plugins/marketplace.json), copies plugin files to plugins/crit/ (project) or ~/.codex/plugins/crit/ (global), enables the plugin in ~/.codex/config.toml, and turns on features.plugins, features.hooks, and features.plugin_hooks.

Plugin source files live in integrations/codex/plugin/crit/. See integrations/codex/README.md for layout and manual setup.

`crit install codex` vs `crit install codex-plugin`

	`crit install codex`	`crit install codex-plugin`
Scope	Skills only (project or global)	Skills + Codex plugin + plan hook
What's installed	`$crit` and `crit-cli` skills under `.agents/skills/`	Same skills, plus `plugins/crit/` (or `~/.codex/plugins/crit/`) with bundled skills and a `Stop` hook
Plan mode	Agent must write the plan to a file before `$crit` works	Hook captures in-chat proposed plans (`<proposed_plan>`) and reviews them automatically

Both approaches give you $crit and the crit-cli skill. Only codex-plugin adds the proposed-plan hook — without it, typing $crit on an in-chat plan (e.g. after choosing "No and stay in Plan Mode") does nothing useful because there is no file path for crit to open.

Disable automatic plan review per shell or globally with export CRIT_PLAN_REVIEW=off. Manual crit plan invocations are unaffected.

Invocation policy

The interactive crit review cycle is wording-gated, not hard-disabled. Skill descriptions and workflow docs tell agents to launch Crit only when the user invokes the platform command (/crit, $crit, /skill:crit, /crit.md, or Windsurf's /crit, as appropriate) or directly asks to use Crit. A normal request to review code, a plan, a diff, a PR, or a page does not count.

crit-cli is intentionally model-discoverable. It teaches agents how to leave and reply to Crit comments, interpret review JSON, share reviews, and synchronize GitHub PR feedback, but it does not start the interactive review cycle.

crit-story is wording-gated like /crit: agents author a chaptered story overview only when the user invokes /crit-story (or the tool equivalent) or directly asks to generate a crit story. After ingest they reconnect with bare crit and continue the same wait → address → next-round cycle. It does not run as part of a normal review request.

The only automatic interactive path is a lifecycle hook immediately after planning mode. The Claude Code plugin, Codex plugin, and Gemini CLI integration retain their existing plan-exit hooks.

What these do

All integrations follow the same pattern:

Pick the review target — current git changes by default, or an explicit file, plan, PR, or commit range when the user names one
Launch Crit — the agent runs the matching crit command to open the review in your browser
Address feedback — after review, the agent reads the review file to find your inline comments and revises the target
Continue the review loop — the agent reruns the printed next-round command until you finish with no unresolved comments

Each integration also teaches the agent about:

crit comment — leave inline review comments programmatically without opening the browser
review file format — how to read comments, resolve them with threaded replies
crit pull/push — sync reviews with GitHub PRs (push supports --event approve|request-changes|comment)