srinitude skills
Portable Agent Skills from one canonical source tree.
This repository is an Agent Plugins 1.0.0 package of portable skills that follow the Agent Skills specification, plus a read-only local MCP server.
What ships
- Canonical skill packages under
skills/.
- Root
plugin.json and mcp.json for Agent Plugins v1 clients.
- Trigger, behavior, failure, recovery, and speed evaluations.
- A bundled, read-only local MCP server.
- Existing native plugin metadata and adapters for client-specific routes.
- Strict JSON Schemas for manifests, reports, benchmarks, and checkpoints.
No client integration contains a second skill body.
GitHub release versions and skill metadata versions are independent. Skills don't depend on each other's versions.
Agent Plugins v1
Compatible clients discover the root plugin.json, every immediate child in skills/, and the root mcp.json. Installation, permissions, and trust remain client-controlled.
See Agent Plugins v1 for the fixed package layout, current client registry, validation evidence, schema provenance, and security boundary. PD-101: that page owns portable-package operating detail, loads for Agent Plugins work, and links back here.
Install
npx skills add srinitude/skills
Use npx skills add srinitude/skills --list to inspect available skills first. The CLI reports anonymous install telemetry to skills.sh unless DISABLE_TELEMETRY=1 is set.
Clone the repository and run mise run bootstrap plus mise run build-mcp when developing the package. Node 24 or newer is required. Mise pins the full local toolchain.
Client support
| Client | Route | Local MCP |
|---|
| MCP-capable Agent Plugins v1 clients | Root portable package | Yes |
| Claude Code | Root plugin and marketplace | Yes |
| Codex | Root plugin and marketplace | Yes |
| ChatGPT | Codex plugin format | Yes |
| Gemini CLI | Root extension | Yes |
| Cursor | Agent Plugins bundle | Yes |
| OpenClaw | Codex-compatible bundle | Yes |
| Hermes Agent | Root Python plugin | Skills only |
| opencode | Project config plus skills install | Yes |
| Continue | Skills CLI adapter | Optional |
| Aider | Read-only config | No |
Claude Code
/plugin marketplace add srinitude/skills
/plugin install srinitude-skills@srinitude-skills
The plugin loads ./skills and .mcp.json from the same checkout.
Codex and ChatGPT
Start Codex with plugin support, open /plugins, add srinitude/skills as a marketplace, and install srinitude-skills.
codex --enable plugins
The same package metadata is the ChatGPT plugin route documented by the Codex plugin format.
Gemini CLI
gemini extensions install https://github.com/srinitude/skills
gemini-extension.json starts the bundled MCP server from the installed extension path.
Cursor
Load the checkout as a local root Agent Plugin. Cursor reads plugin.json, discovers the canonical skills/ tree, and starts the read-only MCP route from mcp.json. The .cursor-plugin/plugin.json manifest keeps Cursor-format compatibility without owning a second skill or MCP definition. This repository does not claim a Cursor Marketplace listing.
OpenClaw
Use the tag-pinned compatible bundle route and cold-discovery check in the OpenClaw adapter note.
Hermes Agent
Use the Skills Hub tap or optional Python plugin route in the Hermes Agent adapter note. Both load the existing SKILL.md path.
opencode
Use the repository installation above. opencode.json starts the local MCP server when running opencode from this checkout.
Continue and Aider
Use the Continue adapter note or Aider adapter note. Both point to the canonical SKILL.md rather than copying it.
Local MCP server
Build it with:
mise run build-mcp
The stdio server exposes six read-only tools:
list_skillssearch_skillsget_skillget_referenceget_eval_manifestvalidate_skill
Paths are confined to the repository skill tree. Absolute paths, traversal, hidden files, nested references, and symlink escapes fail closed. The server has no write tool, telemetry, credentials, or network call.
Validate and evaluate
SKILL_NAME=your-skill-name
mise run validate -- --all --report .artifacts/skill-validation.json
mise run eval -- --skill "$SKILL_NAME" --transport fixture --report ".artifacts/evals/$SKILL_NAME-fixture"
mise run benchmark -- --skill "$SKILL_NAME" --transport fixture --samples 1000 --report ".artifacts/benchmarks/$SKILL_NAME-fixture.json"
Fixture results prove runner behavior only. They are not evidence about a language model. Paid OpenRouter evaluation is a separate post-release step with a frozen model inventory, checkpoint files, a cost estimate, and explicit spend approval.
Use the fixed-route, capped procedure in OpenRouter sweeps.
Run the complete local gate with:
mise run ci
All currently published skills declare metadata.scope: "user" for availability across projects. The factory also supports project scope for availability in one project. Updates retain valid scope, and untouched legacy skills remain readable. The scope and variant workflow covers independent variants and explicit refreshes; the shared-directory integration verifies placement separately.
Contribute and report problems
Read CONTRIBUTING.md before changing a skill or integration. The skills.sh publishing notes record discovery, telemetry, API, and listing-correction rules. Security reports follow SECURITY.md. Other help is covered by SUPPORT.md.
License
MIT
