Plugin Distribution
Sequant is available as a Claude Code plugin — install it directly from Claude Code without npm, and get skills, hooks, and MCP tools with zero configuration.
Prerequisites
Section titled “Prerequisites”- Claude Code — claude.ai/code
- GitHub CLI —
gh auth status(authenticated) - Git — any recent version
- Node.js 22.13+ —
node --version(for MCP server vianpx)
Install the plugin
Section titled “Install the plugin”In Claude Code:
/plugin install sequant@sequant-io/sequantThis loads:
- 17 workflow skills (
/assess,/spec,/exec,/qa,/fullsolve, etc.) - Pre/post-tool hooks (security guardrails, timing, smart tests)
- MCP server (
sequant_status,sequant_run,sequant_logs)
Configure your project
Section titled “Configure your project”/sequant:setupThe setup skill:
- Checks prerequisites (git, gh, node)
- Creates
.sequant/settings.jsonwith defaults - Detects your package manager and dev server URL
- Creates
../worktrees/feature/directory - Copies the constitution template to
.claude/memory/ - Creates
.sequant-manifest.jsonwith stack info
What You Can Do
Section titled “What You Can Do”After setup, all sequant workflows are available:
/assess 123 # Analyze issue, get recommended workflow/fullsolve 123 # End-to-end: spec → exec → qa → PR/spec 123 # Plan implementation/exec 123 # Implement in isolated worktree/qa 123 # Code review and AC validationMCP tools are also available for programmatic access:
sequant_status # Check issue progresssequant_run # Execute workflow phasessequant_logs # Review past run resultsWhat to Expect
Section titled “What to Expect”- Plugin install: Instant — skills and hooks load immediately
/sequant:setup: ~10 seconds — creates config and detects stack- MCP server: Starts automatically via
npx -y sequant@<version> serve(stdio transport), pinned to the plugin’s release version (#793) - First
/fullsolve: 10–30 minutes depending on issue complexity
Plugin vs npm
Section titled “Plugin vs npm”| Capability | Plugin | npm |
|---|---|---|
| Skills (17 slash commands) | Yes | Yes (via sequant init) |
| Hooks (guardrails, timing) | Yes | Yes (via sequant init) |
| MCP tools | Yes (auto) | Yes (via .mcp.json) |
CLI (sequant run, sequant doctor) | No | Yes |
| CI/headless mode | No | Yes |
| TypeScript library exports | No | Yes |
Plugin is for interactive Claude Code users. npm is for power users, CI pipelines, and programmatic access.
Keeping the plugin up to date
Section titled “Keeping the plugin up to date”Plugins do not auto-update. Claude Code pins an install to the version (commit SHA) you installed it at and never moves it on its own — even as new releases ship. An install can silently run months-old skills and hooks while newer versions are available.
To pick up a new release, update the installed plugin and restart Claude Code:
claude plugin update sequant@sequantNote the distinction from refreshing the marketplace listing:
| Command | What it does |
|---|---|
claude plugin update sequant@sequant | Updates the installed plugin to the latest marketplace version (restart Claude Code after). |
/plugin marketplace update sequant-io/sequant | Refreshes only the local marketplace listing — it does not touch any installed plugin. |
Staleness warning
Section titled “Staleness warning”pre-tool.sh compares the running plugin’s version against the local marketplace clone and prints a single warn-only line to stderr when the install falls behind:
sequant plugin v<installed> is stale (marketplace has v<latest>) — run: claude plugin update sequant@sequant, then restart Claude CodeThe check is zero-network (grep/sed/awk only), never blocks the tool call, and is rate-limited to once per day via a stamp file. It only fires for real plugin-cache installs (paths under plugins/cache/) — repo-local dev copies never warn. See Plugin Updates & Versioning for the full versioning and auto-update opt-in details.
Configuration
Section titled “Configuration”After /sequant:setup, edit .sequant/settings.json to customize:
| Setting | Description | Default |
|---|---|---|
run.timeout | Max phase duration (seconds) | 1800 |
run.qualityLoop | Auto-iterate on QA failures | false |
run.maxIterations | Max quality loop iterations | 3 |
run.smartTests | Auto-run tests after edits | true |
run.pmRun | Package manager run command | auto-detected |
run.devUrl | Dev server URL for /test | auto-detected |
agents.parallel | Parallel sub-agent execution | false |
agents.model | Sub-agent model (currently inert — see anthropics/claude-code#43869; kept for forward compatibility) | "haiku" |
MCP Server Details
Section titled “MCP Server Details”The plugin bundles an MCP server that starts automatically. The bundled
.mcp.json pins sequant to the plugin’s release version (#793) — the build
step stamps the concrete version in place of @latest, so an npx reconnect
never re-resolves @latest (which was surfacing as -32000 after releases).
You pick up a new version when you update the plugin:
{ "sequant": { "command": "npx", "args": ["-y", "sequant@2.9.0", "serve"] }}Tools: sequant_status, sequant_run, sequant_logs
Resources: sequant://state, sequant://config
Transport: stdio (default, most compatible)
Log Storage
Section titled “Log Storage”Hook logs are stored in the first location available:
- Plugin users:
${CLAUDE_PLUGIN_DATA}/logs/(persistent across updates) - npm/CI users:
${HOME}/.sequant/logs/(persistent; one sink shared by the main repo and every worktree) - Last resort (no
HOME):${TMPDIR:-/tmp}/— session-scoped, and macOS purges it
Log files: claude-timing.log, claude-hook.log, claude-quality.log, claude-tests.log
pre-tool.sh and post-tool.sh resolve this location with an identical block and share claude-timing.log (START from pre, END from post) and claude-quality.log. Changing the location in one hook without the other splits every START/END pair across two files (#763).
The log directory deliberately lives outside the repo: creating it inside one would make git status --porcelain non-empty, which silently defeats the pre-tool no-changes guard that reads exactly that output.
Troubleshooting
Section titled “Troubleshooting”MCP tools not available after install
Section titled “MCP tools not available after install”Symptoms: Skills work but sequant_status / sequant_run don’t appear.
Solution: Verify Node.js 22.13+ is installed (node --version). The MCP server requires npx which ships with Node.js. Restart Claude Code after installing Node.
/sequant:setup says “gh auth: not authenticated”
Section titled “/sequant:setup says “gh auth: not authenticated””Symptoms: Prerequisites check fails on GitHub CLI.
Solution: Run gh auth login in your terminal, then re-run /sequant:setup.
Settings not being picked up
Section titled “Settings not being picked up”Symptoms: Changed .sequant/settings.json but behavior doesn’t change.
Solution: Verify the file is valid JSON (cat .sequant/settings.json | jq .). Settings are read fresh on each skill invocation — no restart needed.
Plugin install says “not found”
Section titled “Plugin install says “not found””Symptoms: claude plugin install sequant@sequant-io/sequant fails.
Solution: Verify you’re using Claude Code v2.1+. Try the full path: claude plugin install sequant@sequant-io/sequant.
Generated for Issue #476 on 2026-03-27