Agents create, flip and roll back flags
Agents use the same verbs as people: list, get, create, set, target, kill, bulk, revert and eval, through MCP, the flagger CLI or the API, for example flagger create checkout_v2 --summary "New checkout".
Flagger.gg is the feature-flag system built for AI agents: agents create, flip and roll back flags, every change has guardrails and an audit log, and people approve what goes live. Everything marked live works today in the hosted service; the rest is in development.
Agents use the same verbs as people: list, get, create, set, target, kill, bulk, revert and eval, through MCP, the flagger CLI or the API, for example flagger create checkout_v2 --summary "New checkout".
Connect any agent to mcp.flagger.gg with its token, or run flagger mcp on stdio. Agents only see the tools their token allows, and there is no approve tool.
Each agent gets its own token for one project, a list of environments and the actions you pick (read, create, target, rollout, kill, request), with a 90-day expiry. Tokens are stored as hashes and can be revoked any time.
A person turns flags live from the console. Flags that aren't live yet come first, each with a line on what changes for customers, a live switch, bulk changes and the flag's history.
Per project. In auto-live mode agents' changes apply at once. With approvals on, agents can roll out to a cap you set per environment, and anything wider waits for a person to approve the request.
A bad state, a percentage outside 0 to 100 or a target list with nobody on it is rejected. Protected flags refuse agents except for the kill switch, and unknown flags are off, so mistakes fail safe.
off for no one, admins for your admins, targets for named users and workspaces, percent for a sticky share of users, and on for everyone.
A user's bucket comes from a SHA-256 hash of the flag key and user id, so the same user stays in or out as you raise the percentage from 0 to 100.
Target user ids or verified emails (matched without case) and workspace ids. Rules are kept when the state changes.
One call sets a flag off for everyone and keeps its rules. Revert rolls a flag back to its previous state, or to any point in its history.
Every change, by a person, an agent, an admin token or a deploy, is stored with who made it, the agent's run, when, the state and rules before and after, and a note. Agents must give one.
Flags are evaluated in your app against an in-memory ruleset that streams updates and is cached on disk, so a check never makes a network call. Unknown, not yet loaded or errored means off.
Your app registers its flags with a key, description, owner and summary when it starts. The console shows which flags are in code, and a deploy never overwrites a state someone set.
The SDKs download your rules and decide locally. The users you check flags for are never sent to Flagger.gg.
withFlag forces a flag for one test, setFlagOverride for longer, and bothFlagStates runs a test with the flag off and then on, in both SDKs.
Import an app's existing flags with their history, then run the SDK in shadow mode, which compares its answers with your current system before you switch.
Every console view, filter and open flag has its own address, so a reload or a shared link shows the same thing.
Watch error rates, latency and your own business metrics next to each flag's state and rollout.
Step a percentage rollout up on a schedule and switch the flag off automatically if a guarded metric regresses.
Define an audience once, such as beta workspaces or internal staff, reuse it across flags, and target on more attributes.
Split users between variants and compare them on the metrics you pick.
Libraries for more languages beyond Go and Node, with the same off-by-default rules and shared test vectors.
Sign up, ship your first flag, and tell us how your team and its agents work.