Feature flags for AI agents

From setup to a running flag

1. Create the app

   An authorized setup agent can sign up or log in with email and password through the CLI, without a browser. Use an owner-controlled email and pass the password through stdin. The original account login creates the app and saves its one-time keys to a private file.

   Headless signup and login →

2. Give the worker scoped access

   Create a separate token for each worker. Choose read, manage, or admin permission within its account, app, or environment scope. Hand it over through the cloud provider's secret store as SUPERFLAG_TOKEN. The worker does not need the account password or original login token.

   Set up a cloud agent →

3. Create flags and generate types

   A worker with manage permission can create flags, targeting rules, and percentage rollouts in its assigned scope. The CLI exposes a machine-readable command contract and JSON results. Generate TypeScript flag names and value types from config, and check for drift in CI.

   CLI workflows and code generation →

4. Use the flag in your app

   Node, React, and React Native SDKs evaluate flags from cached configuration. Give server code an sdk_ key and client apps a pub_ key for client-visible flags. These evaluation keys are separate from the worker's management token. Run the app and verify both the enabled and disabled paths.

   Integrate your first flag →

Give one worker seven days of access

superflag tokens create \
  --name cloud-my-app-dev \
  --app my-app --env dev \
  --permission manage --expires-in 7 \
  --write-secrets /private/worker.json --json

Run this on the trusted provisioner after creating the app. Use a new file in a private directory. The file receives the token; stdout contains metadata. Store only its token field in the worker's runtime secret. Verify the scope with whoami --json from the worker.

Can a scoped worker create more credentials?

No. Scoped tokens cannot create apps, SDK keys, or other tokens, even with admin permission. The trusted provisioner performs that setup. A worker cannot expand its own access.

Can agents use MCP?

The hosted MCP endpoint at https://superflag.sh/api/mcp is read-only. Local stdio MCP also supports proposals and an approval-gated execution flow. Approval uses separate authority, and a producer cannot approve its own proposal. Direct flag writes use the CLI or HTTP API.