tool-list-mod

<!-- markdownlint-disable MD033 MD041 -->

<img src="docs/banner.png" alt="OrchestKit - Stop explaining your stack. Start shipping." width="100%" />

<!--ork:skills-->110<!--/ork--> skills · <!--ork:agents-->36<!--/ork--> agents · <!--ork:hooks-->169<!--/ork--> hooks

---

<a href="https://orchestkit.yonyon.ai/"><strong>Explore the Docs →</strong></a> · <a href="https://yonyon.ai/go/orchestkit?utm_campaign=readme"><strong>OrchestKit Community →</strong></a><br> _{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.

Surface	Claude Code	Cursor	Codex	pi	Devin	Antigravity
Skills (SKILL.md)	all	all, via the ork plugin	6 (ork-codex pack)	all via pi install, 78 auto-listed	76 of 107, GH-4146	all via workspace .agents/skills (skills.sh); ork:<name> via agy plugin install
Agents	all	all	4 role templates	none	not reported by info	36 via agy plugin install (validated); none via .agents/skills
Hooks	all	none	none	none	none	none of ork's; agy hooks.json is a different schema
Rules	repo convention	14, plugin rules key	AGENTS.md	none	AGENTS.md, always on	AGENTS.md, per-directory
Commands	/ork:<skill>	36 wrappers	$ork-<skill>	/skill:<name>	/ork:<skill>	/<skill-name>; ork:<skill> via plugin install
MCP config	.mcp.json	.cursor/mcp.json	plugin mcp.json	.pi/mcp.json	mcp.json / .mcp.json	~/.gemini/config/mcp_config.json or plugin mcp_config.json; repo .mcp.json not read
Status	shipped	shipped	shipped	shipped	skills only, GH-4146	skills only, measured agy 1.2.6

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 from measured installs plus the ork skills we retain (full why table under Install):

npx skills add yonatangross/orchestkit -s devops-deployment -s responsive-patterns -s architecture-decision-record -s ui-components -s rag-retrieval -s agent-orchestration -s brainstorm -s expect -s auto -s review-pr -s fix-issue -s commit

Full catalog: npx skills add yonatangross/orchestkit (hundreds of SKILL.md files; do not treat that as unique users). Live install totals stay on the skills.sh badge above.

---

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:

Without	With OrchestKit
"Use FastAPI with async SQLAlchemy 2.0..."	"Create an API endpoint" → Done right
"Remember cursor pagination, not offset..."	Agents know your patterns
"Don't commit to main branch..."	Hooks block bad commits
"Run tests before committing..."	/ork:commit runs tests for you

---

What You Get

One unified plugin, everything included.

Component	Details
<!--ork:skills-->110<!--/ork--> Skills	RAG patterns, FastAPI, React 19, testing, security, database design, ML integration — loaded on-demand, zero overhead
<!--ork:agents-->36<!--/ork--> Agents	Specialized personas (backend-architect, frontend-dev, security-auditor) — route tasks to the right expert
<!--ork:hooks-->169<!--/ork--> Hooks	Pre-commit checks, git protection, quality gates, browser safety — ship with confidence

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

Server	Purpose	Required?
Context7	Up-to-date library docs	Prerequisite (22 of 36 agents grant its tools)
Memory	Knowledge graph persistence	Recommended
Sequential Thinking	Structured reasoning for subagents	Recommended
Tavily	Web search and extraction	Optional

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 <!--ork:hooks-->169<!--/ork--> hooks across <!--ork:events-->32<!--/ork--> 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.

What	Destination	Notes
Lifecycle events (session end, PR merged, goal converged, chain phase)	~/.local/state/orchestkit/events.jsonl	Written unconditionally, rotated at 10 MB. ORK_EVENTS_LOG redirects the path (used by the test suite)
Hook metrics: event name, tool name, payload size, duration	same local file	Size-capped metrics only
Prompt text and file contents	Never recorded	Hooks read them to make an allow/deny decision, then discard
Remote sync	Off	No endpoint is compiled in; see below

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. Picked from measured skills.sh installs on our listing plus the ork skills we actually retain:

skill	why
devops-deployment	most-installed cloud skill on our listing
responsive-patterns	most-installed design skill on our listing
architecture-decision-record	writing-bucket install leader
ui-components	design install co-leader
rag-retrieval	SE install leader
agent-orchestration	agent-workflow install leader
brainstorm	our most-retained ork skill
expect	browser verification we run day to day
auto	front-door router we actually invoke
review-pr	ship-loop review we use
fix-issue	debug loop we use
commit	ship-loop commit (has a with/without receipt)

npx skills add yonatangross/orchestkit -s devops-deployment -s responsive-patterns -s architecture-decision-record -s ui-components -s rag-retrieval -s agent-orchestration -s brainstorm -s expect -s auto -s review-pr -s fix-issue -s commit

All skills: npx skills add yonatangross/orchestkit. Live install totals stay on the skills.sh badge above, not hard-coded here.

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 you