AI assistants (MCP)

featflags runs a hosted MCP server, so any AI assistant that connects to remote MCP servers can work with your flags: “turn search-v2 off in production”, “roll new-checkout out to 10%”, “which flags can we delete?”. There’s nothing to install and no key to copy: you add a URL, sign in with your featflags account, and choose what the assistant may do.

https://app.featflags.io/mcp

Connect your assistant

When the assistant first connects, a featflags page opens in your browser: sign in if you aren’t, pick the organisation the assistant works in, and choose whether it gets your own access there or read-only access. Then allow it.

Claude Code. Run this, then /mcp in Claude Code and choose to authenticate:

claude mcp add --transport http featflags https://app.featflags.io/mcp

Claude (claude.ai or the desktop app). Settings → Connectors → Add custom connector, and paste the URL.

Cursor. Settings → MCP → Add new MCP server, or in ~/.cursor/mcp.json:

{ "mcpServers": { "featflags": { "url": "https://app.featflags.io/mcp" } } }

VS Code. Command Palette → MCP: Add Server → HTTP, or in .vscode/mcp.json:

{ "servers": { "featflags": { "type": "http", "url": "https://app.featflags.io/mcp" } } }

Anything else that supports remote MCP servers with OAuth sign-in takes the same URL.

The sign-in page tells you what’s asking. An assistant that registered itself is shown by the name it claims (“an app calling itself Claude”), since any app can call itself anything; one identified by its own website shows that site. When the sign-in returns to an app on your computer, the page says so, because any app on the computer could be listening: only allow a connection you just started.

What it can do

Tool	What it does
list_projects	Your projects and their environments, with which need a comment or approval.
list_flags	A project’s flags and their status in each environment, searched by key, description or status.
get_flag	A flag’s variants and, per environment, whether it’s on, its rules, default and release plan.
evaluate_flag	What a flag serves a given context in an environment, and why.
flag_history	A flag’s changes, newest first.
flag_stats	Its evaluations by variant and hour.
stale_flags	Flags ready to archive, those safe to remove because no scanned code reads them, and others that stopped changing.
flag_code_references	Where code reads a flag, per repository and branch.
create_flag	Creates a flag, off everywhere.
set_flag_enabled	Turns a flag on or off in an environment.
update_targeting	Changes a flag’s rules or default in an environment, percentage rollouts included.
start_release_milestone	Starts, or goes back to, a milestone of the flag’s release plan.
archive_flag	Archives a flag.

Reading tools are marked read-only and changing ones as such, so your assistant can ask you before it changes anything.

Safety

• It acts as you. The assistant has your access in the organisation you picked, or a reader’s if you chose read-only, and never more. If your role shrinks, or you leave the organisation, its access does too.
• Your rules apply. Its changes go through the same checks as yours: where an environment requires a comment, it has to give one; where it requires approval, its change waits for reviewers and it tells you so. Its changes are recorded in history as yours, through the assistant.
• It can’t widen its reach. No tool manages members, API tokens, SDK keys, experiment keys or billing, or shows a key or a secret. Individual targets, uploaded lists, segments, webhooks, triggers and experiments stay in the app.
• It’s easy to disconnect. Your account page lists your Connected apps, with each one’s organisation, access and last use, and a Disconnect button. Owners can also revoke any connection among the organisation’s API tokens. A connection you don’t use for 90 days expires.

featflags doesn’t run an AI model: your assistant’s model decides which tools to call, and featflags only sees those calls.