Flaggr ships agent skills — packaged instructions that teach coding agents (Claude Code, Devin, Cursor, and any agent that reads SKILL.md conventions) how to drive flaggr.dev correctly: evaluation, toggles, rollout safety, SDK wiring, browser analytics, and error and crash reporting.

Agents load the one that matches the task, so asking for "set up error tracking" or "why did LCP get worse" brings in the right instructions without loading the rest.

Install with npx

# Auto-detects .agents/skills, .claude/skills, or .devin/skills in your project
npx @flaggr/skills
 
# Install globally — available in every project
npx @flaggr/skills --global
 
# Explicit target directory
npx @flaggr/skills --target .claude/skills

Preview first: npx @flaggr/skills --list and --dry-run.

Claude Code and Codex plugins

The Flaggr plugin is published to npm as @flaggr/plugin. One install adds these skills and connects the hosted MCP server at https://flaggr.dev/mcp — sign-in is OAuth, so there are no tokens to copy. In Codex it also adds the Flaggr app entry.

Claude Code

Inside Claude Code, add Flaggr's marketplace:

/plugin marketplace add https://flaggr.dev/plugins/marketplace.json

Then install the plugin:

/plugin install flaggr@flaggr

Or from a terminal: npx @flaggr/plugin --claude. It runs the same two steps with the claude CLI, for your user.

Either way needs Node.js 18 or later: Claude Code downloads the plugin with npm, so npm must be on the PATH Claude Code sees (a session started from an IDE or the desktop app may not see an nvm or mise install). No Node.js? Add just the MCP server, without the skills: claude mcp add --transport http --scope user flaggr https://flaggr.dev/mcp.

Then start a new session (or run /reload-plugins), run /mcp, select plugin:flaggr:flaggr (the Flaggr plugin's server) and choose Authenticate.

Codex

npx @flaggr/plugin --codex

The installer needs Node.js 18 or later. It writes a one-plugin Flaggr catalog to ~/.flaggr/codex-marketplace (keep that directory: Codex reads the catalog from there), registers that directory as a Codex marketplace, and runs codex plugin add flaggr@flaggr.

Then open Codex and sign in to Flaggr when prompted. Codex runs the bundled flaggr-setup skill on install and lists the Flaggr app under Explore — choose Pin to sidebar.

npx @flaggr/plugin with no flag installs into every client it finds on your PATH; --dry-run prints the commands without running them. If you already have a marketplace named flaggr (a checkout of the plugin, say), the installer keeps it and installs from it, refreshing it first where the client can; for Codex it then writes nothing to ~/.flaggr.

Updates

• Claude Code: inside Claude Code, choose Update now under /plugin → Installed → flaggr, or run /plugin marketplace update flaggr, which refreshes the catalog and updates the plugins installed from it. From a terminal, run claude plugin update flaggr@flaggr (or npx @flaggr/plugin --claude again); in a terminal, claude plugin marketplace update flaggr only refreshes the catalog. Then run /reload-plugins or restart Claude Code. To get updates automatically, turn on auto-update under /plugin → Marketplaces → flaggr.
• Codex 0.143 or later: automatic. Codex picks up a new version the next time the Codex app or TUI starts.
• Codex 0.131–0.142: re-run npx @flaggr/plugin --codex to refresh the local copy.

Prefer to wire the server yourself? See Connect Claude, Codex & AI Assistants. To run the server locally over stdio with a token instead — a personal access token (fgp_…) for every project you can access, or a project API token (fgr_…) for one — use npx -y @flaggr/mcp-server@latest (see the MCP server guide). It needs version 0.3.0 or later: earlier versions can run shell commands taken from a crafted url passed to the login tool. @latest installs the newest release on npm, so until npm view @flaggr/mcp-server version prints 0.3.0 or later, use the hosted server instead.

Bundled skills

flaggr-setup

Onboarding — runs when the plugin is installed in Codex, and whenever you ask to set up or reconnect Flaggr:

• Connects through the client's native OAuth flow and verifies it with get_connection
• Explains read / write / admin access and what a 403 means
• Walks the setup checklist with get_onboarding_status: project, service, first flag, SDK, browser analytics

flaggr

Core operations — the default skill agents load when you mention flags:

• Evaluation (evaluate_flag, POST /api/flags/evaluate, batch eval)
• Flag CRUD and toggles (toggle_flag, bulk_toggle_flags)
• Health and safety (get_flag_health, check_and_rollback_flag, toggle-impact analysis)
• Auth — token minting, Bearer scopes, public-flag evaluation
• REST quick reference: sparse fieldsets (fields), limit with offset or a cursor, bulk updates with summary, the version-rollback route, and If-None-Match (304), which only the pixel-grid demo list answers

flaggr-rollouts

Progressive delivery workflows — activated when you ask for rollouts, canaries, or kill switches:

• start_canary_rollout — starts a persisted canary rollout plan gated on the flag's pre-canary error rate / p99, shown under Releases
• check_and_rollback_flag — circuit-breaker rollback on error/latency drift
• simulate_flag_targeting — blast-radius estimation before changing rules
• Cohorts, overrides, experiments (their variants are assigned only on the control plane) and bandits (experimental)

The flaggr and flaggr-rollouts skills tell agents that impact analysis, toggle-impact checks and circuit-breaker rollbacks see only the evaluations https://flaggr.dev serves and the exposures reported to it. The data plane (api.flaggr.dev) records none.

flaggr-analytics

Browser analytics (real-user monitoring) — activated when you ask about Web Vitals, page views, conversions or the analytics script:

• Creating a Site (create_site) and adding the a.js tag with its site_… id, the data-* options, Content Security Policy and version pinning
• Tagging events with the page's flag variants: window.__FLAGGR_FLAGS__, which @flaggr/sdk 0.5.0 and later fills in, or by hand for other providers
• Conversions (flaggr:conversion) and remote collection control with the _flaggr_analytics flag (levels, sampling, the attribute floor)
• Reading the results: get_web_vitals, query_browser_segments, get_flag_rum_report for a variant's effect on vitals, errors and conversions, and the /api/analytics/* routes
• Verifying the install (list_recent_browser_events), and what the script and the ingest keep

flaggr-errors

Error and crash reporting, Sentry-style — activated when you ask to capture errors, read a stack trace, triage issues, upload source maps or look at latency:

• Setup per platform: the analytics script's data-track-errors, flaggrErrors() in @flaggr/sdk (browser and Node.js), the Go SDK's errorreport package, @flaggr/react-native for React Native and Expo, and POST /api/ingest/server-events from any language, with which credential each needs
• Triage with list_issues, get_issue (including the flag variants an issue's events carried) and update_issue_status, and how resolved issues regress
• Source map uploads (PUT /api/projects/{id}/sourcemaps, or the CLI's sourcemaps upload), release health and crash-free sessions, and the issue alert conditions
• Performance: transactions and traces (get_performance, get_server_trace)
• Which version of each SDK and of the CLI each integration needs

flaggr-openfeature

SDK integration — activated when wiring Flaggr into application code:

• @flaggr/sdk: createFlaggr, the React and Next.js provider, and the stream / batch / poll update modes
• Flaggr's Go SDK and Python SDK (v0.1.1; Python installs from its GitHub release tag)
• OpenFeature OFREP providers, with the base URL https://api.flaggr.dev/api
• Transport selection — when to use REST eval vs SSE push vs Connect-RPC

Authentication

With the plugin, the hosted connector signs in with OAuth. Without it, skills use a token: a project API token (fgr_…, created in the project's settings) for SDKs and flag evaluation, or a personal access token (fgp_…, Profile → Personal access tokens) for scripts and the local MCP server. API tokens and personal access tokens can't create tokens. For read-only exploration, the public pixel-grid demo service works with no token at all — try it in the API explorer.

See also

• MCP server guide — the underlying tool surface
• API explorer — live Try-It for every endpoint
• Authentication — token types and scopes