Agent Toolkit
Agent Toolkit gives coding agents reusable skills and tools for web research, documentation lookup, read-only .NET quality analysis, skill authoring, and planning across sessions.
Each capability lives in a focused plugin, combining Markdown guidance with scripts, standalone agents, or MCP servers where needed. Install individual skills or choose a plugin for your coding tool. Shared source keeps the content consistent across hosts, while each integration respects the host's capabilities.
Multi-harness support
Choose an installation route for your coding tool. Available components depend on the route and the host's capabilities.
| Host | Installation route | Available components |
|---|---|---|
| Claude Code | Native plugin marketplace | Canonical plugins and skills; optional Claude-specific planning hooks |
| OpenAI Codex CLI | Native plugin marketplace | Canonical plugins and skills |
| Cursor | Native plugin marketplace | Canonical plugins and skills |
| OpenCode | make install-native-opencode | Generated skills, portable agents, and commands with translated tool permissions |
| Antigravity CLI | make install-native-antigravity | Generated plugins containing skills and portable agents; commands become manual skills |
| GitHub Copilot | make install-native-copilot | Generated agent profiles and skills; commands become manually invoked skills |
| Gemini CLI | make install-gemini-cli | Canonical skills and supporting assets |
| Pi | make install-pi | Canonical skills and supporting assets |
The dependency-free skills-only installer also supports OpenCode, Antigravity, and Copilot. Do not mix installation routes in the same scope: generated skill names can duplicate bare-name skills.
See Host Harnesses for the capability matrix, configuration paths, model mappings, tool permissions, and installation limits.
Quick start
Choose a plugin for your task:
| Plugin | Purpose |
|---|---|
| Search | Bounded public web retrieval and cited research through a .NET MCP |
| Dotnet Quality | Read-only dotnet format analysis through a standalone .NET agent |
| Docs Research | Official documentation research and implementation checklists |
| Skill Guide | Design, author, and validate reusable skills |
| Planning | Durable task journals, verified progress, and session recovery |
See the plugin catalog for details and the usage guide for commands and examples.
Claude Code
Run inside Claude Code:
/plugin marketplace add mehdihadeli/agent-toolkit
/plugin install agent-toolkit-planning@agent-toolkit
Choose another registered plugin for a different workflow. Planning hooks are optional and Claude-specific; see the Planning README before enabling them.
Codex CLI
Run these commands in a terminal with a current Codex CLI:
codex plugin marketplace add mehdihadeli/agent-toolkit
codex plugin add agent-toolkit-planning@agent-toolkit
The native Codex manifests expose canonical source skills through the committed
.agents/plugins/marketplace.json; no generation step is required. This
repository does not generate Codex TOML agents or command-derived skills.
Installing a plugin does not start .NET agents or MCP servers. See
Codex setup details for the native installation
route and its boundaries.
Cursor
Import https://github.com/mehdihadeli/agent-toolkit through Cursor's marketplace
import flow, then install a plugin from Customize. Cursor reads the committed
.cursor-plugin/marketplace.json and canonical skills.
See native marketplace setup for host-specific requirements and local testing.
Skills only
Install an individual skill with GitHub CLI's gh skill command or Vercel's
npx skills installer. Select your coding tool when prompted. Both fetch
skills from GitHub without requiring a local clone and install them without
the plugin's agents, commands, or hooks.
gh skill install mehdihadeli/agent-toolkit durable-plans
npx skills add mehdihadeli/agent-toolkit --skill durable-plans
gh skill requires GitHub CLI 2.90 or later. The npx command requires Node.js
and npm. See skills-only installers
for available skills, agent selection, install scopes, and version pinning.
For owned installs from a local clone, the
Python installer remains available.
Install from a clone
For OpenCode, Antigravity CLI, and Copilot native artifacts, clone the repository and install the pinned Python adapter dependency:
git clone https://github.com/mehdihadeli/agent-toolkit.git
cd agent-toolkit
make deps
Choose the target for your host:
make install-native-opencode PLUGIN=planning DRY_RUN=1
make install-native-opencode PLUGIN=planning
# Alternatives: make install-native-antigravity or make install-native-copilot
Make is optional; equivalent Python commands are documented in Host Harnesses. Antigravity filesystem staging may also require native plugin-manager registration. That guide covers project scope, updates, installation checks, and safe removal.
What's inside
| Component | What it provides |
|---|---|
| Five plugins | Independent release boundaries with source, manifests, and usage documentation |
| Skills | Markdown guidance, references, and scripts loaded for relevant tasks |
| .NET MCP | Public web tools exposed over stdio or Streamable HTTP |
| .NET agent | Standalone read-only quality analysis using the Copilot SDK |
| Host adapters | Generated native formats for portable Markdown skills, agents, and commands |
Browse Plugins, Agents, and Agent Skills for details.
How it works
Hosts discover a plugin through its registry and manifests, then consume its registered surfaces. Installing one plugin does not load the entire repository into the agent's context.
plugins/docs-research/
+-- .claude-plugin/plugin.json
+-- .codex-plugin/plugin.json
+-- .cursor-plugin/plugin.json
+-- README.md
`-- skills/docs-research/
`-- SKILL.md # Reusable Markdown guidance
Native registries reference canonical source. Adapters generate ignored host-specific artifacts from portable Markdown surfaces. Commands may become skills, and model choices and tool permissions vary by host. Models inherit host defaults unless explicitly overridden.
Skills and Markdown agents provide instructions. .NET agents run as standalone applications; MCP servers expose tools over stdio or Streamable HTTP when explicitly launched or configured in a host. External providers and credentials are configured through environment variables.
Adapters do not convert .NET applications into Markdown agents or translate MCP configuration, secrets, or Claude hooks. MCP clients connect to servers that users explicitly host or configure.
Source stays inside plugins/<id>/; tests and evaluations live under
tests/<id>/. See Architecture for ownership and runtime
boundaries.
Quality evaluation
Validation covers different kinds of behavior:
- Repository checks validate manifests, marketplace alignment, links, and evaluation coverage.
- Waza provides credential-free trigger coverage for skills and host-native agents.
- Vally evaluates skill activation, prompt contracts, agent behavior, and MCP tool calls.
- xUnit v3 unit and integration tests exercise executable .NET plugins.
- Deterministic script and installer tests check skill tooling and host adapters.
Install development dependencies and run static checks:
mise install
npm ci
mise run validate
mise run vally-lint
mise run waza-eval
Check generated artifacts and run executable tests:
make generate-all
make garden
dotnet test --solution agent-toolkit.slnx
make test-harness
Run prompt evaluations through GitHub Copilot or the optional Claude executor:
mise run vally-eval-copilot-local
npm run build --prefix tools/vally-executor-claude
mise run vally-eval-claude-local
Prompt evaluations require host authentication and provider configuration. The Claude executor runs the same Vally suites through Claude Code. See Plugin Evaluation, Tests, and CI for credentials, focused checks, and local equivalents.
Documentation map
| Guide | Covers |
|---|---|
| Plugin catalog | Plugins by task and their host registrations |
| Agents and skills | Components, conventions, and progressive disclosure |
| Usage and host setup | Commands, installation routes, and capability limits |
| Architecture and packaging | Source ownership, runtimes, adapters, and distribution |
| Evaluation and CLI verification | Prompt checks, real-host recipes, and their limits |
| Tests and CI | Test layout, credentials, and local workflow equivalents |
| Documentation website | Local preview and GitHub Pages deployment |
Preview the documentation website with npm ci and npm run docs:dev.
Contributing: CONTRIBUTING.md. Repository conventions and quality gates: AGENTS.md. See authoring conventions when adding plugin content.
License
This project is licensed under the MIT License.