OrchestKit Docs MCP

107 skills · 36 agents · 171 hooks

Explore the Docs → · OrchestKit Community →
_{Skill browser, demo gallery, setup wizard}

Contents

• Quick Start
• Why OrchestKit?
• What You Get
• Key Commands
• Configuration
• What OrchestKit observes
• Install
• FAQ
• Development
• What's New
• Community

Quick Start

Pick the host you actually use. Claude Code is the full plugin (skills + agents + hooks). Cursor gets the same ork plugin minus Claude hook scripts. skills.sh is skills only — start with the 12 below, not the whole catalog.

Host support matrix

Measured 2026-09-08 on pi 0.85, Codex CLI and cursor-agent. Details, commands and the lane model: OrchestKit on pi, Codex and Cursor. Antigravity measured 2026-09-18 on agy 1.2.6; full evidence in docs/audits/agy-host-support-2026-09-18.md.

Claude Code

/plugin marketplace add yonatangross/orchestkit
/plugin install ork

Then /ork:setup. The wizard scans the repo, recommends skills, and writes MCP config.

CLI equivalent: claude plugin marketplace add yonatangross/orchestkit && claude plugin install ork@orchestkit.

Cursor

Settings → Plugins / marketplaces → add yonatangross/orchestkit → enable ork → open a new chat. Same plugin Claude Code installs, not a five-skill fork. See Install → Cursor.

skills.sh (Cursor, Codex, OpenCode, …)

Starter 12 — doctor, setup, explore, implement, verify, review-pr, commit, expect, assess, brainstorm, create-pr, remember:

npx skills add yonatangross/orchestkit -s doctor -s setup -s explore -s implement -s verify -s review-pr -s commit -s expect -s assess -s brainstorm -s create-pr -s remember

The skill named implement is the implement workflow (/ork:implement in Claude Code). There is no ork-implement on the Claude plugin; Codex uses $ork-implement after the Codex pack is installed. Full catalog: npx skills add yonatangross/orchestkit (hundreds of SKILL.md files — do not treat that as unique users).

Why OrchestKit?

Every Claude Code session starts from zero. You explain your stack, patterns, preferences—again and again.

OrchestKit gives Claude persistent knowledge of production patterns that work automatically:

What You Get

One unified plugin, everything included.

All available in a single /plugin install ork. Skills load on-demand. Hooks work automatically.

Browse everything in the Docs →

Key Commands

/ork:auto         # Front door: describe a goal, it routes to the right skill
/ork:setup        # Personalized onboarding wizard
/ork:implement    # Full-stack implementation with parallel agents
/ork:expect       # Diff-aware AI browser testing
/ork:review-pr    # PR review with parallel agents
/ork:verify       # Multi-agent validation
/ork:commit       # Conventional commit with pre-checks
/ork:explore      # Analyze unfamiliar codebase
/ork:remember     # Save to persistent memory
/ork:doctor       # Health check

Configuration

/ork:setup detects your stack, recommends MCP servers, and writes the configuration for you.

Recommended MCP Servers

Set "alwaysLoad": true on the first three in your .mcp.json. It skips the per-skill tool probe and shaves ~150ms off cold starts.

Context7 is a prerequisite, and ork does not ship it. 22 agents grant mcp__context7__* in their frontmatter, but .mcp.json is user-owned and project-scoped, so the grant refers to a server you add. Skip it and those agents answer from training data with no error raised. The recommended entry is the hosted HTTP server, which costs no local process:

"context7": {
  "type": "http",
  "url": "https://mcp.context7.com/mcp"
}

Free tier: 1,000 requests, public repos, no account. Context7 Pro ($10 per seat per month) raises that to 5,000 per seat and parses private repos; add "headers": { "Authorization": "Bearer ${CONTEXT7_API_KEY}" } and export the ctx7sk- key. Add the header only once the variable is exported: with it unset the unexpanded literal is sent as the token and every query fails, and it does not fall back to the anonymous free tier, so the keyless entry above is strictly better than a header with no key behind it. The legacy stdio transport (npx -y @upstash/context7-mcp@4.0.2) is the fallback when the hosted endpoint is unreachable, but it spawns one child process per Claude Code session, so the fan-out scales with how many sessions you keep open.

Customizing skills

Skills install as files on your disk, but don't hand-edit the installed copy — it gets overwritten on update and silently diverges from the canonical playbook. The supported ways to extend (user-level skills, project skills, upstream PRs, or disabling a bundled skill) are in docs/extending-skills.md.

What OrchestKit observes

OrchestKit is a quality-gate plugin, so its hooks are the product rather than an add-on. This section states plainly what they see, where it goes, and how to turn each piece off.

Scope: broad and intentional. OrchestKit registers 171 hooks across 32 lifecycle events, including SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, and Stop. They are not gated to a particular framework or project type, because the gates they enforce (secret-write blocking, protected-file guards, git safety, file-size limits, agent status protocol) apply to any codebase. If you only want gates on some projects, enable the plugin per-project rather than globally.

Where data goes: a local file on your own disk.

There is deliberately no global kill switch for the local write, because the gates depend on that state (the git-safety and chain-staleness hooks read their own prior events). To stop it entirely, disable the plugin. Individual noisy hooks have their own opt-outs: ORK_DISABLE_DEBT_TRACKER, ORK_DISABLE_WORKTREE_VERIFIER, ORK_DISABLE_COORDINATION_METRICS, ORK_NO_NOTIFY, ORK_NO_STALE_SWEEP, and ORCHESTKIT_SKIP_SLOW_HOOKS among others.

Network access is opt-in and unset by default. There is no hardcoded remote host anywhere in the shipped hook bundles (grep -o 'https\?://' plugins/ork/hooks/dist/*.mjs returns nothing). An outbound call happens only if you configure a destination yourself, via one of:

• ORCHESTKIT_HOOK_URL + ORCHESTKIT_HOOK_TOKEN, which enable the manual hooks/bin/telemetry-sync.mjs CLI. It POSTs your local JSONL to your own endpoint. No hook ever invokes it; you run it by hand.
• ORK_HQ_TELEMETRY_URL, which points the telemetry HTTP sink at your own collector.
• ORK_HQ_TELEMETRY_USE_HQ_API=1 together with HQ_API_URL, the same sink aimed at a self-hosted HQ API.
• ORK_SESSION_CATEGORY_PROVIDER=jev (or shadow) together with the TypeSafe key variable ORK_TYPESAFE_API_KEY, a second classifier for the session work category. The session-identity hook already asks a local claude -p --model haiku process for a title and a category; with both variables set it also asks TypeSafe's Jev model (api.typesafe.ai, model pinned to jev-1.13.0) one typed Choice over the same eight categories and the same criteria text. This is the one exception to the "no hardcoded host" note above, and it is dormant unless both variables are set. What leaves your machine: the git branch name and the first 600 characters of the session's first prompt, sent to a third-party processor under its own data policy. What changes in jev mode: when Jev answers at confidence 0.8 or above, its category decides the session color (held out on 150 sessions, that band is 93.5% correct); below 0.8, or on any error, haiku's category decides as before. The title and emoji always come from haiku. In shadow mode nothing you see changes; Jev is only logged beside haiku. To turn it on locally, in the shell that launches claude: export ORK_TYPESAFE_API_KEY="$(<your secret manager> ...)" and export ORK_SESSION_CATEGORY_PROVIDER=jev. Both outcomes are logged once per session in the hook log, as category jev: haiku=<cat> jev=<cat> agree=<bool> confidence=<n> decided_by=jev|haiku threshold=0.8 latency_ms=<n>, and as session-identity.shadow.json with the same fields (labels, decision, timing; never prompt text) next to the raw answer session-identity.jev.json in the session data directory. Cost of opting in: the first prompt of a session waits for the call to settle, 1.1 to 1.7 s measured from a fresh hook process (the eval's 320 ms was a warm connection), 3 s at most before it gives up.
• ORK_ROUTE_JEV=shadow (or steer) together with ORK_TYPESAFE_API_KEY, the Jev routing seam for /ork:auto (#4233). Off by default. When set, every build-shaped prompt (the same test the once-per-session executor reminder uses: an imperative build or fix verb, 40 characters or more, no explicit /plugin:skill) is sent to the same TypeSafe endpoint and model as one typed Choice over nineteen route classes plus four side judgments (needs a worktree, needs a browser, mutation risk, needs the operator). What leaves your machine: the first 1,500 characters of the prompt after a redactor has replaced secrets, emails, Israeli phone numbers, nine digit ids, op:// references and the names of directories under clients/ with [SECRET], [EMAIL], [PHONE] and [CLIENT], plus the repository basename. The serialized request is scanned again before it leaves; anything that survives redaction refuses the call. In shadow mode nothing you see changes; the verdict is logged. In steer mode, at confidence ORK_ROUTE_JEV_FLOOR (default 0.5) or above, the one-line executor reminder names the executor the class maps to (route: dev_fix -> /ork:fix-issue (conf 0.83)); the model still reads the /ork:auto table and says whether it agrees. Fallbacks are fail open: no key, timeout, non-2xx, malformed answer, below the floor, or the daily token budget (ORK_ROUTE_JEV_DAILY_TOKENS, default 2,000,000) spent all mean the existing path runs untouched; a 402 or 429 switches the seam off for 24 hours. One log line per prompt, route jev: intent=<class> conf=<n> top3=<a:p,b:p,c:p> worktree=<n> browser=<n> mutation=<n> operator=<n> floor=<n> decided_by=jev|table|off|budget|egress latency_ms=<n> input_tokens=<n> redacted=<n>, and one record per prompt in jev-route.jsonl next to session-identity.jev.json (labels, timings and a sha256 of the redacted text; never prompt text). Offline replay: node scripts/eval/route-check.mjs --jev and node scripts/eval/jev-route-score.mjs. The vendor's agent skill is a peer plugin installed from its own marketplace, never a copied directory: claude plugin marketplace add typesafe-ai/skills then claude plugin install typesafe@typesafe-ai; /ork:doctor reports the installed version against the marketplace and prints the update commands.

The sink returns early when the URL or the token is missing, and telemetry-sync.mjs prints No ORCHESTKIT_HOOK_URL or TOKEN configured. Nothing to sync. then exits 0. There is no analytics ping, no crash reporter, and no feature-flag fetch.

What OrchestKit never reads. No OS keychain lookups, no ~/.aws/credentials, no SSH private keys, no browser cookie or login stores, no clipboard. The one place secret-shaped paths appear in the source is plugins/ork/hooks/dist/pretool.mjs, where id_rsa, .pem, .env, and credentials.json form a blocklist that stops Claude writing to them. That code denies access; it does not read those files.

Third-party MCP servers are recommendations, not bundled dependencies. The plugin ships no .mcp.json and declares no mcpServers. The table under Configuration is advisory, and /ork:setup asks before writing anything.

Install

/plugin install ork

No tiering. No version confusion. Just one powerful plugin.

Not on Claude Code? Pull a starter 12 into any agent (Cursor, Codex, OpenCode, …) via skills.sh — do not install the whole firehose on day one:

npx skills add yonatangross/orchestkit -s doctor -s setup -s explore -s implement -s verify -s review-pr -s commit -s expect -s assess -s brainstorm -s create-pr -s remember

All skills: npx skills add yonatangross/orchestkit. The implement skill is implement, not ork-implement.

Cursor

Cursor loads Agent Plugins and Cursor plugins. This repo already ships the Agent Plugins manifest at plugins/ork/plugin.json. Add the GitHub repo as a Cursor marketplace (Settings → yonatangross/orchestkit), enable ork, then open a new chat. That is the same plugin Claude Code installs, not a five-skill fork.

It also ships 14 rules under the plugin's rules key, generated from src/rules/ and src/shared/rules/. They are agent-fetched, so a rule costs context only when its description matches the task.

Claude hook scripts are not registered for Cursor: they depend on ${CLAUDE_PLUGIN_ROOT} (orchestkit#293, closed). Cursor enforcement for HQ repos stays in the consuming project's .cursor/hooks.json.

"Include third-party Plugins" can leak SKILL.md from ~/.claude/plugins. That is not an install. Proof is the ork plugin id plus the full skill catalog.

Codex

Codex uses its own plugin format, skill picker, and standalone role configuration. Add OrchestKit's Codex marketplace, then install the small portable workflow pack:

codex plugin marketplace add yonatangross/orchestkit --ref main --sparse .agents/plugins --sparse plugins/ork-codex
codex plugin add ork-codex@orchestkit-codex

Restart Codex after installation. Invoke a workflow explicitly with $ork-brainstorm, $ork-explore, $ork-implement, $ork-assess, $ork-verify, or $ork-review-pr; their narrow descriptions also let Codex select the relevant workflow automatically.

The plugin intentionally ships roles as templates because Codex loads custom roles from ~/.codex/agents/, not from a plugin manifest. From an OrchestKit checkout, run this one-time, non-overwriting install:

plugins/ork-codex/scripts/install-codex-roles.sh ~/.codex/agents

It installs ork_explorer, ork_implementer, ork_reviewer, and ork_verifier; restart Codex before spawning them.

Unattended runs: the ork-mech profile

For mechanical work (renames, bumps, codemods, sweeps that end in a diff), install the shipped profile and run codex exec against it:

plugins/ork-codex/scripts/install-codex-profile.sh ~/.codex
codex exec --profile ork-mech "<task>" </dev/null

The profile is a FILE, not a snippet you paste into config.toml. Measured on codex-cli 0.153.4: --profile <name> layers $CODEX_HOME/<name>.config.toml over the base config, and a legacy [profiles.<name>] table left inside config.toml makes the same flag a hard config-load error. The installer refuses to run next to that table, and refuses to overwrite a profile you already have.

What it sets, as codex exec prints it in its own header:

approval: never
sandbox: workspace-write [workdir, /tmp, $TMPDIR] (network access enabled)
reasoning effort: high

It deliberately does not pin a model (pass -m) and does not use --dangerously-bypass-approvals-and-sandbox, which drops the sandbox entirely. Two contracts a TOML file cannot express, so they stay on the command line:

• </dev/null. codex exec reads stdin even when a prompt argument is given. An inherited open pipe blocks the run with Reading additional input from stdin... and no timeout.

• --add-dir inside a git worktree. The writable roots are [workdir, /tmp, $TMPDIR]. A linked worktree's git common dir sits outside the workdir, so the first commit dies on index.lock. Add it:

  codex exec --profile ork-mech \
    --add-dir "$(git rev-parse --path-format=absolute --git-common-dir)" \
    "<task>" </dev/null
  

Keeping the install current

ref main in the marketplace source is a cached snapshot, not a tracker. On the 2026-09-08 audit machine codex plugin list showed 10.0.0-beta.5 while main was three releases ahead. After an OrchestKit release, update and check:

codex plugin update
codex plugin list | grep ork-codex          # installed version
jq -r .version plugins/ork-codex/.codex-plugin/plugin.json   # what main ships

Documentation lookup (context7)

The plugin ships a context7 MCP server in its own manifest (mcpServers in .codex-plugin/plugin.json, defined in mcp.json), so installing the plugin registers it. Confirm with codex mcp get context7. It is scoped to the only two tools context7 exposes, resolve-library-id and query-docs, and it uses the hosted HTTP transport rather than an npx stdio child, so it costs no extra process per Codex session.

Export a key before starting Codex. The plugin references the variable name and never stores the value, so no token is written to ~/.codex/config.toml:

export CONTEXT7_API_KEY_CODEX="<your-context7-api-key>"

Put that in your shell profile so every Codex session inherits it. Get the key from your own context7 account and keep the value out of the repository. If you store it in a secret manager, substitute your own vault and item names (with the 1Password CLI the reference is op://<vault>/<item>/credential), and cache the resolved value instead of re-reading the vault in every shell: each raw read is a separate unlock prompt.

Two behaviors worth knowing:

• A server you already define yourself under [mcp_servers.context7] in ~/.codex/config.toml wins, and the plugin's definition is ignored entirely (including its tool scoping). That is intentional: your own configuration is never overridden. Remove your entry if you want the plugin's.
• Without a valid key the server still connects and still lists its tools. Only a real call fails, with Invalid API key. A successful connection is therefore not proof of authentication.

pi

pi (0.85) reads the same SKILL.md format, and the repo now carries a pi manifest, so the package installs directly:

pi install git:github.com/yonatangross/orchestkit

That registers every skill. Add -l to write .pi/settings.json in the project instead of your user settings. Pointing pi at a checkout still works and needs no install (pi --skill ./plugins/ork/skills).

The 29 skills marked disable-model-invocation stay reachable only as /skill:<name>. Two measured caveats: --no-builtin-tools hides every skill (pi lists skills only when a file-reading tool is enabled), and pi -p blocks on an open stdin, so headless runs need </dev/null.

MCP servers for pi come from .pi/mcp.json, then .mcp.json, then ~/.config/mcp/mcp.json. Copy the shipped template to get the recommended servers with a read-only includeTools allowlist per server:

cp .pi/mcp.json.example .pi/mcp.json

Full detail and the tracking epic: OrchestKit on pi, Codex and Cursor.

Devin

devin plugins install yonatangross/orchestkit

Devin checks a plugin root for a manifest in this order: .devin-plugin/plugin.json, then .claude-plugin/plugin.json, then root plugin.json (plugins reference). The repo root only had the last one, the Agent Plugins manifest, which fixes skills at a root skills/ directory this repo does not have, so a bare install used to list 0 of 107 skills, 0 hooks, and only the AGENTS.md rule (GH-4146). A root level .devin-plugin/plugin.json now maps skills to ./plugins/ork/skills, the same built tree Claude Code and Cursor already read.

devin plugins info ork lists 76 of 107 skills after that change (measured on Devin CLI 3000.10.27). The other 31, auto, verify, help, brainstorm, and similar router or workflow skills, declare a triggers: frontmatter key shaped as an object (keywords, examples, anti-triggers). Devin's own triggers field expects a flat [user, model] list, and the shape mismatch drops the whole skill instead of warning. Renaming that key touches three other readers (scripts/eval/eval-coverage.sh, tests/skills/triggering/test-trigger-keywords.sh, and the SKILL_ONLY allowlist in tests/plugins/test-command-frontmatter-passthrough.sh), so it stays tracked on GH-4146 instead of folded into this fix. Installing the subpath directly, devin plugins install yonatangross/orchestkit#plugins/ork, reaches the same 76 skills plus the plugin's 36 custom subagents, which neither path surfaces through devin plugins info. That subpath is what this repo's own .claude-plugin/marketplace.json already points Claude Code at.

OrchestKit's plugin hooks work under Claude Code only today; none of the 171 fire under Devin. Two reasons. Location: OrchestKit ships hooks at plugins/ork/hooks/hooks.json, while Devin's own plugin format reads a bare hooks.json at the plugin root instead. Shape and tool names: OrchestKit's file matches Claude Code tool names such as Bash and Write|Edit and expands ${CLAUDE_PLUGIN_ROOT} in command args, both Claude Code specific. Devin's own hooks.v1.json format matches its own tool names instead, exec, write, edit and friends (see lifecycle hooks), inside the same {matcher, hooks:[{type, command}]} event map shape. devin plugins info confirms Hooks (none) even once skills resolve. PreToolUse and PostToolUse read the closest to what OrchestKit's own pretool and posttool hooks already do, and are the first candidates for a future Devin hook manifest once the tool name mapping is written deliberately instead of guessed.

Pin a release instead of tracking main: release-please tags every release as v10.0.0-beta.N. The single argument devin plugins install <source> command has no ref pinning syntax of its own; pin through a requiredPlugins entry (in this repo's own .devin/config.json, or a personal, org, or enterprise manifest) instead:

{
  "requiredPlugins": [
    { "source": "github", "repo": "yonatangross/orchestkit", "ref": "v10.0.0-beta.30" }
  ]
}

Antigravity

npx skills add yonatangross/orchestkit -s doctor -s setup -s explore -s implement -s verify -s review-pr -s commit -s expect -s assess -s brainstorm -s create-pr -s remember

Antigravity (agy, Google's agentic CLI) reads Agent Skills from a workspace .agents/skills directory, so the shared skills.sh line above is the install path and Antigravity is already in that installer's universal target list (measured on agy 1.2.6; full evidence in docs/audits/agy-host-support-2026-09-18.md). Verify from the repo root with agy -p "/skills" --add-dir "$PWD": print mode only registers a workspace from an absolute --add-dir, while the interactive TUI takes the launch directory as the workspace and needs no flag.

Two honest gaps. npx skills add -g lands in ~/.agents/skills, which agy does not read; user scope lives at ~/.gemini/config/skills (or ~/.gemini/skills). And none of ork's 171 hooks port: agy's hooks.json uses its own event names and payload shape, and agy plugin validate plugins/ork reports hooks skipped because ork keeps them at hooks/hooks.json, not the plugin-root hooks.json agy looks for.

A fuller install exists but needs a local checkout: agy plugin validate <checkout>/plugins/ork reports skills: 108 processed, agents: 36 processed, and agy plugin install <checkout>/plugins/ork copies them into ~/.gemini/config/plugins/ork/ namespaced ork:<name>, which matches the /ork:<skill> spelling. Plugin-bundled mcp_config.json also works (marker-verified end to end), but a repo-level .mcp.json is never read.

FAQ

/plugin list
/plugin uninstall ork && /plugin install ork

Development

npm run build      # Build plugins from src/
npm test           # Run all tests

Edit src/ and manifests/, never plugins/ (generated).

See CONTRIBUTING.md for details.

What's New

v10.0.0-beta.70 · 2026-09-20

• cc: raise the CC support floor 2.1.251 to 2.1.277 (#4305)

v10.0.0-beta.69 · 2026-09-20

• ci: skip the playground gate on test-only diffs (#4303)

v10.0.0-beta.68 · 2026-09-20

• jev: add shared shadow foundation and passive Codex adapter (#4286)

v10.0.0-beta.67 · 2026-09-20

• expect: jev live per-field decisions in form steps (#4284)

v10.0.0-beta.66 · 2026-09-20

• glyph: add a triage page template with a parsable DECIDE pick (#4267)
• jev: persist paired shadow decisions across local seams (#4278)

v10.0.0-beta.65 · 2026-09-20

• jev: add conservative weak outcome sidecars (#4281)
• docs-site: stop the picker repeating the hero install command (#4293)

v10.0.0-beta.64 · 2026-09-20

• site: search autocomplete suggestions in docs search (#4258)
• hooks: never emit hookSpecificOutput without hookEventName (#4285)

v10.0.0-beta.63 · 2026-09-19

• eval: document Langfuse Python SDK v4 and fail zero-item runs (#4272)

See CHANGELOG.md for the full release history.

Community

Join the Building with AI community for AI dev tips, OrchestKit support, and connecting with other builders:

Names and audiences match what yonyon.ai renders, so the two surfaces cannot drift. Every link resolves through yonyon.ai/go/*, so a rotated invite never needs a README change and no raw invite is published here.

Who builds this

OrchestKit is built and maintained by Yonatan Gross — Yonyon AI, an AI consulting practice. It is the toolkit extracted from real client work, not a side project: the patterns here are the ones that survived shipping.

It stays MIT and free. Nothing is gated, and none of the below changes that.

Working out where AI actually fits in your business? The AI readiness audit is a free assessment that maps your workflows and returns a prioritized report — the same diagnostic that opens a consulting engagement.

Want the toolkit running properly in your team? Setup, configuration, and a working agent loop tailored to your stack is something I do as a fixed-scope engagement. Start a discussion or reach out through the community.

Security policy and reporting: SECURITY.md.