Kerf MCP server
Connect Kerf to an AI assistant so it can read flags, launch tests and summarize results.
The Kerf MCP server exposes your project to any assistant that speaks the Model Context Protocol. An assistant can list flags, read experiment results and, if you allow it, change rollout percentages.
Install
npx @kerf/mcp-server --token $KERF_API_TOKENConfigure your client
{
"mcpServers": {
"kerf": {
"command": "npx",
"args": ["@kerf/mcp-server"],
"env": { "KERF_API_TOKEN": "kerf_..." }
}
}
}What it can do
- list_flags and get_flag for current state.
- get_experiment_results for lifts, intervals and guardrails.
- draft_readout to write a summary for a finished test.
- set_rollout to change a percentage, only if write access is on.
Permissions
Tokens are read only unless you grant the write scope. Even with write access, the server will not change a flag that has a guardrail breach open, and every change is logged with the assistant's session id.
Example prompts
- Which experiments ended this week and what did we decide?
- Is the new checkout flag safe to move to 50%?
- Write the readout for short-address-form.
Good practice
Use a dedicated token per assistant, scoped to the projects it needs. Review the audit log after a new assistant is connected.