Start here
Installation
Desktop, native CLI, package managers and supported platforms.
CANONICAL SOURCEView this guide in GitHub ↗
Installation
This guide describes the primary consumer install flow for agent-toolkit, then lists advanced methods when you need marketplace plugins, skills-only delivery, or manual file copies.
Prerequisites
The product CLI is the native V binary (GitHub Releases). uvx agent-toolkit-cli / PyPI still wrap that binary (ADR-021). Index: docs/v/README.md. Contributor build: docs/HOW_TO_DEVELOP_V.md.
- Python 3.10 or later — optional unless you use the PyPI trampoline (
uv/uvx) - At least one supported AI coding assistant:
- Claude Code (Anthropic)
- Cursor
- OpenCode
- GitHub Copilot (VS Code or JetBrains + extension)
- Windsurf (Codeium)
- Pi Coding Agent
- Muse Code (Meta)
Optional but recommended:
- uv — preferred way to run
uvxand manage the CLI - gh (GitHub CLI) — required by forge skills (
gh-fix-ci,github-cli-workflow, etc.) - jq — used by several loop templates for JSON processing
- node / npm — for MCP server installation and
npx skills - git + bash — only needed for git-clone or install-script methods below
Swarms — agent-toolkit swarm prerequisites
Swarms are backend-neutral: the orchestration engine is filesystem-based, no cloud required. You need a UI backend and a runner:
- UI backend (one required):
- Herdr — recommended. Rich workspace/tabs UI:
brew install herdrorcurl -fsSL https://herdr.dev/install.sh | sh. Verifyherdr --versionandherdr integration install opencode. See SWARM_HERDR.md. tmux— portable fallback, works over SSH/headless Linux. Installtmuxvia your package manager (brew install tmux,apt install tmux, etc.). Swarm uses an isolated server/socket per runagent-toolkit-swarm-<run-id>and never mutates your normal tmux sessions. See SWARM_TMUX.md.- Use
--backend auto(Herdr → tmux fallback) or explicitly--backend herdr/--backend tmux/--backend headless.swarm doctorreportsherdr available,tmux available, versions, andopencode integration installed/outdated.
- Herdr — recommended. Rich workspace/tabs UI:
- Runner (one required — provides the LLM):
- OpenCode — primary recommended runner (
opencode modelsto listprovider/model). Alternatives:claude,codex,cursor-agent,copilot,muse. Discover viaagent-toolkit swarm models --runner opencodeoragent-toolkit swarm runners. See SWARM_MODELS_AND_COSTS.md. - Offline/fake demo: no runner or Herdr needed — use
--runner skeleton(always available).agent-toolkit swarm start --runner skeleton --backend tmux --dry-run "demo task"previews without creating state; omit--dry-runfor a real local run.
- OpenCode — primary recommended runner (
- Git: required for worktree isolation. Each writing role gets
worktrees/<role>on branchagent-toolkit-swarm/<run-id>/<role>; code moves via validated full 40-char SHAs. - Agentic-workstation auto-provision: if you use agentic-workstation, set
agent_swarms.enabled=trueto install tmux + Herdr + integrations automatically.
Check everything at once:
agent-toolkit swarm doctor
agent-toolkit swarm doctor --json
agent-toolkit swarm backends --json
agent-toolkit swarm runners --json
Primary install (recommended)
Graphical Desktop app
Download the latest Agent Toolkit Desktop release for Linux, macOS, or Windows. The Electron package includes the canonical backend, so a separate CLI install, source checkout, or developer toolchain is not required for normal Desktop use. Start with the Desktop product guide for onboarding, People, projects, Library, Operations, and terminal workflows.
The CLI install below is for terminal-first setup, scripting, CI, and headless workflows. Desktop users can perform their basic workspace setup from the GUI.
Support matrix: the single source is
docs/TRUST.md#Installation channels— GitHub Releases (canonical artifact), PyPI, npm, Homebrew, AUR, GHCR container, Claude/Cursor marketplaces, Agent Plugins artifacts. The canonical artifact is the native V binary from a GitHub Release; PyPI/npm/marketplaces are distribution adapters, Homebrew/AUR are downstream packages that fetch the canonical artifact. V is canonical, Python is a thin launcher.
One command auto-detects your AI tools and deploys the right profiles. Pick one channel — all install the same V CLI.
# Homebrew
brew tap ulises-jeremias/homebrew-tap && brew install agent-toolkit
# AUR (native V; not the Python AUR package)
yay -S agent-toolkit-bin
# GitHub Release — download agent-toolkit-<os>-<arch> + SHA256SUMS from
# https://github.com/ulises-jeremias/agent-toolkit/releases/latest
# PyPI launcher (execs bundled V)
uv tool install 'agent-toolkit-cli>=1.11.0'
# npm
npm i -g agent-toolkit-cli
agent-toolkit install
agent-toolkit doctor
From a git checkout the canonical implementation is the V binary (#555, HOW_TO_DEVELOP_V.md):
./make.vsh install-cli # <prefix>/bin/agent-toolkit; default ~/.local/bin
# custom prefix: ./make.vsh install-cli --prefix=/usr/local
agent-toolkit doctor --json
See docs/v/archive/cutover.md and docs/v/archive/rollback.md. PyPI/uvx ships a thin Python trampoline over the V binary (ADR-021; python-fallback.md). Do not retag empty v1.10.0.
Install options
# Specific tools only
agent-toolkit install --tools claude-code,cursor
# Preview changes without writing files
agent-toolkit install --dry-run
# Overwrite existing toolkit-managed files
agent-toolkit install --force
Verification
After the primary install:
agent-toolkit doctor
# Optional: shell completions (bash / zsh / fish / PowerShell)
agent-toolkit completion bash >> ~/.bashrc
Open your AI tool and ask: “What skills do you have available?” — responses should reflect the agent-toolkit skill set.
To validate skill definitions from a git checkout:
./scripts/validate-skills.vsh
Advanced install methods
Use these when the primary CLI flow does not fit your environment.
Claude Code plugin marketplace
Native plugins for Claude Code only:
/plugin marketplace add ulises-jeremias/agent-toolkit
/plugin install agent-toolkit-core@agent-toolkit
/plugin install agent-toolkit-agents@agent-toolkit
/plugin install agent-toolkit-forge@agent-toolkit
Cursor plugins (IDE + Agent CLI)
Native plugins from .cursor-plugin/marketplace.json:
agent-toolkit-core, agent-toolkit-agents, agent-toolkit-forge.
Cursor IDE
- Open Customize in the sidebar and choose Plugins → Add → From GitHub Repository.
- Enter
https://github.com/ulises-jeremias/agent-toolkit. - Install the plugins you need (
agent-toolkit-coreis the baseline) at user or project scope. See the current Cursor plugin guide.
User-scoped installs sync to Cursor Agent CLI sessions automatically.
Cursor Agent CLI
# Open Cursor Agent and use its current plugin install flow.
# See https://cursor.com/docs/plugins for the current CLI syntax and scope options.
cursor-agent
Load a local plugin directory for one session:
cursor-agent --plugin-dir ./plugins/agent-toolkit-core
cursor-agent --plugin-dir ./plugins/agent-toolkit-agents
cursor-agent --plugin-dir ./plugins/agent-toolkit-forge
Local / offline
mkdir -p ~/.cursor/plugins/local
ln -s "$(pwd)/plugins/agent-toolkit-core" ~/.cursor/plugins/local/agent-toolkit-core
ln -s "$(pwd)/plugins/agent-toolkit-agents" ~/.cursor/plugins/local/agent-toolkit-agents
ln -s "$(pwd)/plugins/agent-toolkit-forge" ~/.cursor/plugins/local/agent-toolkit-forge
See Cursor plugins docs.
npx skills (skills only)
Installs skills via the Agent Skills standard. Does not deploy agents, loops, or full profiles.
npx skills add ulises-jeremias/agent-toolkit -g
Homebrew / AUR
brew tap ulises-jeremias/homebrew-tap && brew install agent-toolkit
yay -S agent-toolkit-bin # Arch Linux (AUR) — GitHub Release V binary
npm i -g agent-toolkit-cli # npm optionalDependencies platform packages @1.11.0
Git clone (from-source CLI)
For offline installs or pinning a specific commit, build/install the V CLI from the checkout
(ADR-007 removed the deprecated scripts/install.sh
wrapper):
git clone https://github.com/ulises-jeremias/agent-toolkit ~/.agent-toolkit
cd ~/.agent-toolkit
./make.vsh install-cli # → ~/.local/bin/agent-toolkit (or --prefix=…)
agent-toolkit install
# Or via the PyPI trampoline against the same checkout:
# uvx --from agent-toolkit-cli --from ~/.agent-toolkit agent-toolkit install
Manual install
Copy profiles yourself when you need full control over paths. Clone the repo first:
git clone https://github.com/ulises-jeremias/agent-toolkit ~/.agent-toolkit
Claude Code
mkdir -p ~/.claude/agents
cp ~/.agent-toolkit/profiles/claude-code/CLAUDE.md ~/.claude/CLAUDE.md
cp ~/.agent-toolkit/profiles/claude-code/settings.json ~/.claude/settings.json
cp -r ~/.agent-toolkit/profiles/claude-code/agents/. ~/.claude/agents/
Restart Claude Code. Agents are available via @agent-name (e.g. @code-reviewer).
Project-level (overrides global):
cd /path/to/your/project
mkdir -p .claude/agents
cp ~/.agent-toolkit/profiles/claude-code/CLAUDE.md .claude/CLAUDE.md
cp ~/.agent-toolkit/profiles/claude-code/settings.json .claude/settings.json
Cursor
mkdir -p ~/.cursor/rules
cp -r ~/.agent-toolkit/profiles/cursor/rules/. ~/.cursor/rules/
Per-project: copy to .cursor/rules/ inside your repo instead.
OpenCode
mkdir -p ~/.config/opencode/agents
cp ~/.agent-toolkit/profiles/opencode/opencode.json ~/.config/opencode/opencode.json
cp -r ~/.agent-toolkit/profiles/opencode/agents/. ~/.config/opencode/agents/
GitHub Copilot
Per-project, committed to the repository:
cd /path/to/your/project
mkdir -p .github
cp ~/.agent-toolkit/profiles/copilot/copilot-instructions.md .github/copilot-instructions.md
Windsurf
WINDSURF_DIR="${HOME}/.codeium/windsurf"
[ -d "$WINDSURF_DIR" ] || WINDSURF_DIR="${HOME}/.windsurf"
mkdir -p "${WINDSURF_DIR}/rules" "${WINDSURF_DIR}/memories"
cp -r ~/.agent-toolkit/profiles/windsurf/rules/. "${WINDSURF_DIR}/rules/"
cp ~/.agent-toolkit/profiles/windsurf/memories/global_rules.md "${WINDSURF_DIR}/memories/global_rules.md"
Pi Coding Agent
mkdir -p ~/.pi/agent/skills
cp -r ~/.agent-toolkit/profiles/pi/skills/. ~/.pi/agent/skills/
MCP setup
MCP gives your AI tool access to external services (GitHub, Slack, Linear, etc.). See MCP.md for per-tool config locations and provider setup.
Staying up to date
Match the channel you installed:
brew upgrade agent-toolkit
# AUR
yay -Syu agent-toolkit-bin
# PyPI launcher
uv tool upgrade agent-toolkit-cli
# npm
npm update -g agent-toolkit-cli
# GitHub Release: download the new binary + SHA256SUMS from /releases/latest
agent-toolkit install --force
From a git checkout:
cd ~/.agent-toolkit && git pull && ./make.vsh install-cli && agent-toolkit install --force
Back up customized profile files before --force. See MIGRATION.md when
switching from profile-copy installs to marketplace plugins.
Data packaging and resolution
The product CLI is the native V binary. Homebrew and AUR consume GitHub Release assets (distribution/, ADR-023/024). The PyPI launcher wheel still bundles capability data; resolution order for that path is ADR-005 / ADR-015.
Troubleshooting
Claude Code: skills not loading
head -5 ~/.claude/CLAUDE.md
Project-level .claude/CLAUDE.md overrides global. Restart Claude Code after changes.
Cursor: rules not appearing
- Confirm
.mdcfiles are in~/.cursor/rules/(global) or.cursor/rules/(project) - Verify YAML frontmatter (
---,description:, closing---) - Restart Cursor
Windsurf: rules not loading
Try ~/.windsurf/ if ~/.codeium/windsurf/ does not exist for your version.
OpenCode: agents not available
ls ~/.config/opencode/agents/
Restart OpenCode after adding agent files.
MCP servers not connecting
- Confirm the server binary is on
$PATH - Confirm env vars (e.g.
GITHUB_TOKEN) are set in the shell your AI tool uses - Check MCP logs in your tool for connection errors
validate-skills.vsh fails
| Error | Fix |
|---|---|
Missing SKILL.md |
Add SKILL.md to the skill directory |
Missing frontmatter name |
Add name: to the --- block in SKILL.md |
Missing frontmatter description |
Add description: to the --- block |
| Secret pattern detected | Remove the credential; use ${ENV_VAR} placeholders |
Related guides
| Guide | Description |
|---|---|
| TARGETS.md | Supported compile targets and capability matrix |
| MIGRATION.md | Move from profile-copy to native plugins |
| MCP.md | MCP provider setup |
See also: TROUBLESHOOTING.md for doctor error recipes.