Connect Claude Code, Claude, Codex, ChatGPT, Cursor or any MCP client to Flaggr in five steps — hosted MCP server, OAuth approval, live verification, and a guided in-chat setup app.
Last updated October 9, 2026
Connect Claude, Codex & AI Assistants
Flaggr runs a hosted Model Context Protocol server at https://flaggr.dev/mcp. Connect it once and your assistant can set up projects, create and roll out flags, investigate incidents and administer access — scoped to the projects and access level you approve.
The fastest path is the interactive guide at flaggr.dev/connect. It has a screen for each step below, copy-ready commands for your client, and a live check that confirms the connection the moment it lands.
1. Choose your assistant
| Assistant | Where it runs | In-chat Flaggr app |
|---|---|---|
| Claude Code | Terminal, VS Code, JetBrains | Setup checklist via tools |
| Claude | claude.ai, Desktop, mobile | Interactive app in the chat |
| Codex | CLI, IDE extension, desktop app | App in Explore, pin to the sidebar |
| ChatGPT | Developer mode connectors | Interactive app in the chat |
| Cursor | Agent and chat | Tools |
| VS Code | Copilot agent mode | Tools |
| Any MCP client | Streamable HTTP + OAuth | Depends on the client |
2. Add Flaggr
There's nothing to paste except the server URL — no API keys. Sign-in happens on Flaggr's approval page in the next step.
claude mcp add --transport http --scope user flaggr https://flaggr.dev/mcpStart Claude Code, run /mcp, select flaggr and choose Authenticate.
The Flaggr plugin (@flaggr/plugin on npm) bundles the same hosted server with agent skills — including a setup skill (Codex runs it on install) — and, in Codex, the app navigation entry.
Claude Code: inside Claude Code, run these one at a time:
/plugin marketplace add https://flaggr.dev/plugins/marketplace.json/plugin install flaggr@flaggr
From a terminal, npx @flaggr/plugin --claude does the same. Either way needs Node.js 18+ with npm on the PATH Claude Code sees, because Claude Code downloads the plugin with npm; without Node.js, use the claude mcp add command above (the server without the skills). Then run /mcp and authenticate plugin:flaggr:flaggr, the Flaggr plugin's server. To update inside Claude Code, choose Update now under /plugin → Installed → flaggr (or run /plugin marketplace update flaggr); from a terminal, run claude plugin update flaggr@flaggr. Then run /reload-plugins or restart Claude Code. In a terminal, claude plugin marketplace update flaggr only refreshes the catalog.
Codex: run npx @flaggr/plugin --codex (needs Node.js 18+). Codex 0.143 or later installs the plugin from npm and picks up a new version the next time the Codex app or TUI starts. Codex 0.131–0.142 get a local copy instead; run the command again to update it.
See Agent Skills & Plugins for details.
3. Approve access
Your assistant opens Flaggr's approval page. Sign in if needed, then choose:
- Projects in scope. The connection can never reach a project you didn't select. No projects yet? Name your first one right on the page. It goes into the organization you pick there (your own organization by default).
- Access level. Your role in each project still applies on top.
Approving a connection needs a verified email address (403 "Verify your email address first" otherwise). The consent page shows how. Choose Send verification email, open the link, then I've verified — retry, which approves the same projects and permissions. GitHub sign-ins need this once.
| Access | What the assistant can do |
|---|---|
| Read your flags (always on) | Evaluate flags; list projects, services and flags; read analytics, browser vitals, rollouts, incidents and the audit log |
| Manage flags & rollouts | Create, toggle and target flags; run rollout plans and experiments; assign, investigate and resolve alert incidents |
| Administer projects & members (opt-in) | Invite and remove members, change roles (except owner), manage teams and flag owners, API tokens, environments and organization settings |
Deleting a project or an organization, and ownership changes (granting, changing or removing the owner role), are never available to a connection — they stay in the dashboard. Organization tools work only in organizations that own at least one of the projects you selected.
4. Confirm the connection
On flaggr.dev/connect, signed in, the Confirm connection step polls for new connections and turns green the moment your assistant finishes the token exchange — showing the client, access level, projects in scope and when it was last used.
From your assistant, ask:
"Which Flaggr account am I connected as?"
It calls get_connection, which reports the server, signed-in account, access level and projects in scope.
5. Start building
Ask "Open Flaggr". In Claude, Codex and ChatGPT this opens the Flaggr app inside the conversation: a setup checklist that walks through each stage with its own screen, then becomes your project home.
| Stage | What the screen does |
|---|---|
| Connect | Shows the server, account, access level and projects for this connection |
| Project | Switch between consented projects, or ask the assistant to create one (admin) |
| Service | Create a service — one per deployable: web, api, worker |
| First flag | Create a boolean kill switch, off by default |
| SDK | Copy-ready React, Node, Python and Go snippets for your service and flag — or ask a coding agent to wire it in. Completes once Flaggr has recorded an evaluation in the last 7 days: one served by https://flaggr.dev, such as the assistant's evaluate_flag, or an exposure reported to it. The snippets evaluate on api.flaggr.dev, which records none |
| Browser analytics (optional) | Add the analytics script so rollouts can gate on LCP, INP and CLS |
Every change the app makes re-checks the connection first; if the server, account or access level changes underneath it, the view stops rather than writing to the wrong place. Buttons that send prompts or share context only act when you click them.
In clients without MCP Apps, ask "What's left to set up in Flaggr?" — get_onboarding_status returns the same checklist with a ready-made prompt for the next step.
Prompts to try
- "Create a boolean flag new-checkout in my web service, off by default."
- "Add the Flaggr SDK to this repo and evaluate new-checkout."
- "Roll out new-checkout to 10% with LCP and error-rate guardrails."
- "Show open alert incidents and assign the checkout one to me."
- "Who owns the flags in the payments team, and which are stale?"
- "Invite alex@acme.com to this project as a member." (admin)
Manage and revoke connections
flaggr.dev/connect lists every connected assistant with its access level, projects and last use; Disconnect revokes it immediately. Tokens are hashed at rest and checked on every request. To change access, disconnect and connect again with different choices.
Troubleshooting
| Symptom | Fix |
|---|---|
401 / "Authentication required" | Re-run the sign-in step for your client (/mcp → Authenticate, codex mcp login flaggr, or Connect in settings) |
| "connector token is read-only" | Reconnect and allow Manage flags & rollouts |
| "lacks the admin scope" | Reconnect and allow Administer projects & members |
| "not scoped to this project" | Reconnect and select that project on the approval page |
| "not scoped to this organization" | Organization tools need a selected project in that organization — reconnect and select one |
| "ownership changes are not available to connectors" | Change owners in the dashboard |
| Flaggr missing from Codex navigation | Open Explore, find Flaggr and Pin to sidebar; start a new chat to refresh discovery |
Opening https://flaggr.dev/mcp in a browser shows this guide | Expected — the URL is for MCP clients |
Next steps
- MCP Server reference — every tool, resource and prompt
- Agent Skills & Plugins — skills for Claude Code and Codex
- Roles & Permissions — how project and org roles combine with connector access