█████╗ ██████╗ ███████╗███╗ ██╗████████╗
██╔══██╗██╔════╝ ██╔════╝████╗ ██║╚══██╔══╝
███████║██║ ███╗█████╗ ██╔██╗ ██║ ██║
██╔══██║██║ ██║██╔══╝ ██║╚██╗██║ ██║
██║ ██║╚██████╔╝███████╗██║ ╚████║ ██║
╚═╝ ╚═╝ ╚═════╝ ╚══════╝╚═╝ ╚═══╝ ╚═╝
████████╗ ██████╗ ██████╗ ██╗ ██╗ ██╗██╗████████╗
╚══██╔══╝██╔═══██╗██╔═══██╗██║ ██║ ██╔╝██║╚══██╔══╝
██║ ██║ ██║██║ ██║██║ █████╔╝ ██║ ██║
██║ ██║ ██║██║ ██║██║ ██╔═██╗ ██║ ██║
██║ ╚██████╔╝╚██████╔╝███████╗██║ ██╗██║ ██║
╚═╝ ╚═════╝ ╚═════╝ ╚══════╝╚═╝ ╚═╝╚═╝ ╚═╝
Curated plugins for coding agents and developer tools
Agent Toolkit is a fast open-source setup, repo intelligence, and enforcement CLI for coding agents and developer tools, plus plugin marketplace metadata and an agent catalog.
@harryy/agent-toolkitis the Bun package users run withbunx.- The hot path is a Rust native binary for fast repo scanning and checks.
- Supported surfaces include Claude Code, Codex, Gemini CLI, Cursor, VS Code Copilot, Windsurf, Antigravity, OpenCode, Junie, and AGENTS.md-aware tools.
- Each plugin repo can carry shared
AGENTS.mdinstructions,.agents/agents.jsonintegration settings, native extension metadata, and generated repo intelligence.
+------------------------------------------------------+
| agent-toolkit |
| (marketplace) |
| |
| +-------------------+ +----------------------+ |
| | dotagent | | vibe-pilot | |
| | 21 skills | | 4 skills | |
| | 20 agents | | kanban autopilot | |
| +-------------------+ +----------------------+ |
| +-------------------+ |
| | web-smith | |
| | 2 skills | |
| | site cloning | |
| +-------------------+ |
+------------------------------------------------------+
For shared teams, make the repo carry the agent contract:
bunx @harryy/agent-toolkit repo setuprepo setup bootstraps the repo, writes .agents/intel/, pins the DotAgent rules into root AGENTS.md, vendors DotAgent skills into .agents/skills/, copies role-profile and command references under .agents/dotagent/, runs agents sync --path ., and checks the repo. Commit those repo files so new contributors get the same rules and skills from git clone. Later runs use the lock's DotAgent version and git revision as the normal no-op gate; use repo setup --force only when you need to recopy the same snapshot. Use repo setup --skip-sync when @agents-dev/cli is not available.
Install, update, or remove personal defaults outside an agentized repo:
bunx @harryy/agent-toolkit setup --dry-run
bunx @harryy/agent-toolkit setup --yes
bunx @harryy/agent-toolkit update
bunx @harryy/agent-toolkit teardown --dry-run
bunx @harryy/agent-toolkit teardown --yesEnd users need Bun only. Rust is not required for normal use because the published package ships prebuilt native binaries.
setup installs the shared rules plugin, preserves existing user rules, and inserts managed Agent Toolkit blocks where supported. This is a personal convenience layer, not the team source of truth. update pulls the latest DotAgent source and reapplies the packaged personal rules without an extra prompt. teardown removes only managed Agent Toolkit blocks and registrations:
- Codex:
~/.codex/AGENTS.md - Claude Code:
~/.claude/CLAUDE.md - Gemini CLI: links the Gemini extension when
geminiis installed - DotAgent source cache:
~/.agent-toolkit/plugins/dotagent
Useful setup flags:
bunx @harryy/agent-toolkit setup --dry-run
bunx @harryy/agent-toolkit setup --yes
bunx @harryy/agent-toolkit update
bunx @harryy/agent-toolkit teardown --dry-run
bunx @harryy/agent-toolkit teardown --yes
bunx @harryy/agent-toolkit setup --all --yes
bunx @harryy/agent-toolkit setup --skip-gemini --yes
bunx @harryy/agent-toolkit teardown --skip-gemini --yes
bunx @harryy/agent-toolkit teardown --keep-source --yesRepo setup:
bunx @harryy/agent-toolkit repo setup
bunx @harryy/agent-toolkit repo setup --skip-sync
bunx @harryy/agent-toolkit repo setup --forceFleet setup:
bunx @harryy/agent-toolkit fleet scan ~/Desktop ~/Code
bunx @harryy/agent-toolkit fleet migrate ~/Desktop ~/CodeRun agents sync explicitly when you want generated local tool config:
bunx @harryy/agent-toolkit repo sync --check
bunx @harryy/agent-toolkit fleet sync --check ~/Desktop ~/CodeAgent Toolkit uses shared repo instructions first, then tool-specific install surfaces where a tool needs them.
Use the native plugin marketplace when installing into Claude Code.
claude plugin marketplace add harryy2510/agent-toolkit
claude plugin install dotagent@agent-toolkit
claude plugin install vibe-pilot@agent-toolkit
claude plugin install web-smith@agent-toolkitUpdate:
claude plugin marketplace update agent-toolkit
claude plugin update dotagent@agent-toolkit
claude plugin update vibe-pilot@agent-toolkit
claude plugin update web-smith@agent-toolkitGemini uses local extensions. Global setup links the extension automatically when the gemini CLI is detected:
bunx @harryy/agent-toolkit setup --yesFor plugin development or a manual install, link the plugin package:
gemini extensions link /path/to/dotagent/plugins/dotagent/gemini-extension --consent
gemini extensions link /path/to/vibe-pilot/plugins/vibe-pilot --consent
gemini extensions link /path/to/web-smith/plugins/web-smith --consentEach Gemini package contains gemini-extension.json plus a GEMINI.md context file.
Codex reads repo instructions from AGENTS.md. For teams, prefer the repo-scoped DotAgent path so rules and skills are committed with the project. Top-level setup can still install managed shared rules into ~/.codex/AGENTS.md for personal defaults.
bunx @harryy/agent-toolkit repo setupCodex plugin metadata lives beside each plugin:
dotagent/plugins/dotagent/.codex-plugin/plugin.json
vibe-pilot/plugins/vibe-pilot/.codex-plugin/plugin.json
web-smith/plugins/web-smith/.codex-plugin/plugin.json
Each plugin repo also includes .agents/plugins/marketplace.json for Codex-compatible local plugin catalogs.
| Agent/tool | Support path |
|---|---|
| AGENTS.md-aware agents | Read committed AGENTS.md directly. |
| Cursor | Run agents sync --path . to generate local Cursor config. |
| VS Code Copilot | Run agents sync --path . to generate local VS Code MCP config. |
| Windsurf | Run agents sync --path .; user-level MCP sync stays local. |
| Antigravity | Run agents sync --path .; global sync is disabled by default. |
| OpenCode | Run agents sync --path . to generate local opencode config. |
| Junie | Run agents sync --path . to generate local Junie MCP config. |
Agent Toolkit builds repo intelligence, repo maps, and quality enforcement from one command surface.
bunx @harryy/agent-toolkit repo intelThis writes repo intelligence files:
.agents/intel/summary.md
.agents/intel/overview.md
.agents/intel/tasks.md
.agents/intel/tooling.md
.agents/intel/routes.md
.agents/intel/api.md
.agents/intel/components.md
.agents/intel/data.md
.agents/intel/database.md
.agents/intel/graph.md
.agents/intel/impact.md
.agents/intel/boundaries.md
.agents/intel/imports.md
.agents/intel/calls.md
.agents/intel/dependencies.md
.agents/intel/symbols.md
.agents/intel/files.md
.agents/intel/env.md
.agents/intel/testing.md
.agents/intel/repo.json
The generated wiki gives agents a deterministic repo map with task read paths, OXC-backed JS/TS AST extraction, pg_query-backed SQL parsing, framework signals, route/API/component/data surfaces, ordered static database reduction, runtime boundaries, env usage, import graph hotspots, change-impact plans, local import adjacency, call-site indexes, external dependency usage, exported symbols, full file inventory, package scripts, and follow-up guidance for deeper exploration.
repo setup also adds an idempotent managed block to root AGENTS.md that tells agents to start at .agents/intel/summary.md before broad exploration. Existing repo instructions are preserved, and .agents/intel/ is generated repo intelligence that may be committed in migrated repos. repo migrate remains an alias for the same repo setup path.
bunx @harryy/agent-toolkit repo checkThe check command enforces:
AGENTS.mdand.agents/agents.jsonexist.AGENTS.mdpoints agents to.agents/intel/summary.mdbefore broad exploration.scripts/agent-checkand committed repo hooks exist for local and CI enforcement. Existing.husky/or.vite-hooks/layouts are respected; new Vite+ repos default to.vite-hooks/, and other repos default to Husky.husky/.AGENTS.mdrecords the active hook manager so Vite+ repo guidance overrides generic DotAgent Husky defaults.- New JavaScript-platform source uses TypeScript, not
.jsor.jsx. oxlint --type-aware --type-checkreplacestscchecks.oxlintreplaces ESLint.oxfmtreplaces Prettier.- Deslop-style checks catch debug statements, placeholders, empty catches, and likely hardcoded secrets.
- Local/private agent state such as
.agents/local.jsonand.agents/generated/is blocked when tracked by git; migrated repo intel and tool adapters such as.agents/intel/,CLAUDE.md,.codex/,.gemini/, and.cursor/are allowed to be committed. - Bootstrapped repo hooks enforce Conventional Commit messages.
- In Supabase projects,
repo check --stagedruns database lint for staged Supabase SQL/config changes, preferring a project script such asdb:lintorlint:dbbefore falling back tosupabase db lint --local --schema public --level warning --fail-on warning. Add a project script when a repo needs to lint extra app schemas.
Use the migration commands to make this standard repeatable:
agent-toolkit repo setup
agent-toolkit fleet bootstrap ~/Desktop ~/Code
agent-toolkit fleet migrate ~/Desktop ~/CodeUse sync commands when @agents-dev/cli is installed:
agent-toolkit repo sync --check
agent-toolkit fleet sync --check ~/Desktop ~/Codepackages:
@harryy/agent-toolkit
bin/agent-toolkit.ts Bun TypeScript wrapper
src/*.ts package runtime glue
crates:
agent-toolkit-core Rust library for scans, checks, hooks, setup
agent-toolkit-cli Rust native binary
The Bun wrapper runs the native binary when present and falls back to cargo run during local development.
Published packages include prebuilt binaries at:
bin/native/darwin-arm64/agent-toolkit
bin/native/darwin-x64/agent-toolkit
bin/native/linux-x64/agent-toolkit
bin/native/linux-arm64/agent-toolkit
Rust is only required for developing Agent Toolkit itself or building release artifacts.
bun test
cargo test
bun run build:native
bun bin/agent-toolkit.ts repo checkbun run build:native reuses bin/native/<target> when the cached artifact manifest matches the current native build inputs. Set AGENT_TOOLKIT_BUILD_NATIVE_FORCE=1 to force a rebuild.
Do not use JavaScript source files, tsc, ESLint, or Prettier in this repo. Use TypeScript with Bun, oxlint --type-aware --type-check, and oxfmt.
| Plugin | Description | What You Get | Plugin ID |
|---|---|---|---|
| dotagent | Code conventions, agents, and tooling for all projects | 21 skills, 20 agents, 6 commands | dotagent@agent-toolkit |
| vibe-pilot | AI kanban autopilot for classify, triage, and status reports | 4 skills, 3 commands | vibe-pilot@agent-toolkit |
| web-smith | Website cloning and static site optimization | 2 skills, headless audit workflow | web-smith@agent-toolkit |
Code conventions, 20 specialist agents, 21 skills, and tooling for all projects.
+------------------------------------------------------------------+
| dotagent |
| "Load a skill, not your whole brain." |
+------------------------------------------------------------------+
| |
| SKILLS (21) AGENTS (20) COMMANDS (6) |
| ---------------- ---------------- ---------------- |
| ui engineer /setup |
| forms-rhf-zod tester /update |
| react-query-mutative product /uninstall |
| zustand-x-ui-state orchestrator /repo-map |
| tanstack-start-cf reviewer /skill-lint |
| supabase-auth-data debugger /deslop |
| cloudflare security | |
| project-setup performance | |
| scaffold ...and 12 more | |
| toolchain | |
| repo-intelligence | |
| agent-routing | |
| react-best-practices | |
| conventions | |
| shadcn | |
| supabase-postgres | |
| vite | |
| repo-map | |
| deslop | |
| testing | |
| debugging | |
| |
+------------------------------------------------------------------+
| Category | Skills | What They Cover |
|---|---|---|
| UI & Components | ui, shadcn, forms-rhf-zod |
Tailwind v4, shadcn/base-ui, CVA, dark mode, react-hook-form + zod |
| State & Data | zustand-x-ui-state, react-query-mutative |
zustand-x v6, React Query, optimistic updates, key factories |
| Frameworks | tanstack-start-cloudflare, vite, cloudflare |
Routing, server functions, SSR, Workers, Wrangler config |
| Backend | supabase-auth-data, supabase-postgres-best-practices |
3 Supabase clients, auth flow, RLS, migrations, query optimization |
| Quality | react-best-practices, conventions, deslop, testing, debugging |
57 perf rules, oxlint/oxfmt enforcement, AI slop detection, Vitest + Playwright, root-cause workflows |
| Tooling | project-setup, scaffold, toolchain, repo-intelligence, agent-routing, repo-map |
DX tooling, fullstack scaffolding, codebase symbol indexing, agent routing |
20 specialist agents covering engineering, testing, product management, orchestration, code review, debugging, security analysis, performance optimization, and more. Each agent has a focused role and knows when to defer to another.
Full documentation: github.com/harryy2510/dotagent
Zero-touch kanban autopilot. Discovers repos, classifies tasks, triages complex work, picks what to build, and launches AI workspaces.
+------------------------------------------------------------------+
| vibe-pilot |
| "Set it and forget it." |
+------------------------------------------------------------------+
| |
| SKILLS (4) COMMANDS (3) |
| ---------------- ---------------- |
| classify /classify |
| triage /triage |
| implement /status-report |
| status-report |
| |
+------------------------------------------------------------------+
| |
| HOW IT WORKS |
| ---------------------------------------------------------------- |
| |
| Discover repos --> Classify tasks --> Triage complex work |
| | | |
| v v |
| Pick next task <-- Launch workspace <-- Break into steps |
| | | |
| v v |
| Implement --------> Commit & update ----> Next cycle |
| |
+------------------------------------------------------------------+
| Skill | What It Does |
|---|---|
| classify | Scans repos and classifies kanban tasks by complexity, domain, and priority |
| triage | Breaks complex tasks into actionable subtasks with dependencies |
| implement | Picks the next task, launches an AI workspace, and builds it |
| status-report | Generates progress reports across all tracked projects |
Full documentation: github.com/harryy2510/vibe-pilot
Clone and optimize websites into Lighthouse-focused Astro static builds.
| Skill | What It Does |
|---|---|
| website-cloner | Scrapes a source site, extracts design tokens, scaffolds Astro, rebuilds pages, and hands off to optimization |
| site-optimizer | Audits and fixes performance, accessibility, SEO, images, fonts, and static-site delivery |
Full documentation: github.com/harryy2510/web-smith
Each plugin repo includes AGENTS.md and .agents/agents.json for the agents CLI:
agents sync --path .
agents watch --path .Team repos can also carry DotAgent as a pinned project dependency:
bunx @harryy/agent-toolkit repo setupThis updates the default DotAgent source cache, commits the DotAgent instruction block in AGENTS.md, copies DotAgent skills into .agents/skills/, keeps role-profile and command references under .agents/dotagent/, and records the snapshot in .agents/dotagent.lock.json. Re-running the command is a no-op while the locked DotAgent version and revision match the refreshed source. Use repo dotagent only when you want to refresh just the DotAgent snapshot. Generated tool files stay local and are ignored by git unless the repo intentionally uses commit-generated sync output.
For repos bootstrapped by Agent Toolkit, scripts/agent-check runs agent-toolkit repo check. Set AGENT_TOOLKIT_SYNC_CHECK=1 when you also want hooks or CI to run agents sync --path . --check. Repo hooks call the same wrapper, so contributors do not need a global Agent Toolkit install.
Gemini CLI native extension wrappers live in plugin packages as gemini-extension/ where needed.
Hariom Sharma, github.com/harryy2510
MIT