Flag docs

Flag is a feature-flag service: you put code behind flags, then decide who sees what without deploying. It is built for AI agents: an MCP server lets your agent read and change flags with the scopes you grant, and its changes to critical environments become change requests people approve. Flag is metered at cost (pricing) and is in an invite-only beta: sign-ups are closed, so you join through an invite from your org.

Start here

• Getting started: sign in, create a flag and an SDK key, and read the flag from code.
• SDKs: the TypeScript SDK (Node, Bun, Deno, browsers, React) with OpenFeature providers, and OFREP for any other OpenFeature SDK.
• REST API: everything the dashboard does, with dry runs and semantic patches.
• AI & MCP and QuickStart: connect Claude Code, claude.ai, Claude Desktop or any MCP client.

The main ideas

Organizations, projects and environments. An organization has members (owner, admin, member or viewer), projects, and in each project environments such as development and production. Every flag exists in every environment of its project, with its own targeting in each. Members, invites and API keys live under Settings.

Critical environments and approvals. Mark an environment critical and people must type its key to change it, and changes made with an API key or by an agent become change requests. Turn on require approval and every targeting change there becomes a change request that needs reviewers.

Flags. A flag has a kind (boolean, string, number or JSON) and up to 64 variations. In each environment it is on or off. When off, it serves its off variation. When on, a context goes through prerequisites (other flags that must serve a given variation), individual targets (keys that always get a variation), rules (conditions on context attributes or segments, each serving a variation or a percentage rollout), and finally the default rule.

Contexts. What a flag is evaluated for: a JSON object with a kind (default user), a key and any attributes, such as {"kind": "user", "key": "u-123", "email": "ada@acme.com", "plan": "pro"}. Several kinds at once make a multi context: {"kind": "multi", "user": {...}, "org": {...}}. Rules can target any kind.

Segments. Reusable groups of contexts per environment: listed keys of any kind, and rules (optionally a percentage of the contexts matching them). Flag rules use them with the is in segment condition.

Semantic patches. Every change, from the dashboard, the API or an agent, is a list of instructions such as {"kind": "turnFlagOn"} or {"kind": "addRule", ...}. Each can be tried as a dry run first, and is written to the flag’s history and the org’s audit log with who made it.

Schedules. A schedule applies targeting instructions at a set time; where approval is required it becomes a timed change request.

One host

Everything is at https://flag.nightroll.app: the dashboard (/app/), the REST API (/api/v1), the MCP server (/mcp), and the data plane SDKs talk to (/sdk/v1, /ofrep/v1).

Reference

• Evaluation spec and SDK protocol: exactly how flags evaluate and how SDKs talk to Flag, with conformance vectors.
• Security, Privacy and Terms.
• Every page is also Markdown at /docs/<page>.md, listed in /llms.txt.

This page as Markdown: /docs/index.md · All docs for agents: /llms.txt