FeatBit MCP is an MCP server that helps both people and AI agents use FeatBit’s feature flag service more easily. It’s a continuously built service that must follow best practices, and we’re exploring whether an AGI‑driven approach can make that a reality. We now include the featbit-skills project in the same loop to validate it with identical inputs and expected outputs.
Technologies we use include:
- Claude Code for code generation, debugging, and testing
- Ralph Loop for autonomous agent framework
- Feature Flag for release control
- Aspire for logging and debugging
Here's how it will work:
- Build a feed queue, where each item includes:
- New documentation feeds for the MCP server
- Prebuilt validation prompts
- Prebuilt input questions and expected document responses
- Take the earliest item in the feed queue and generate code:
- Use Claude Code to generate code
- All changes must be controlled by feature flags
- Test the code to confirm the FeatBit MCP server can be built and deployed:
- Use Aspire for logging and debugging
- Validate results using the Step 1 prompts to confirm the MCP server and featbit-skills outputs meet expectations:
- Use Claude Code to run test cases
- Use an eval service to assess outcomes
- If results are good, release. Otherwise, debug and return to Step 2 with Ralph Loop until expectations are met:
- Feature flags remain for safe rollback and AGI version control
- Only one git branch is allowed for pushes
flowchart TD
A[Step 1: Build feed queue]
Q{More items in queue?}
A --> Q
Q -- Yes: Next item --> B
subgraph "Ralph Loop (Steps 2-5)"
B[Step 2: Generate code]
C[Step 3: Test build & deploy]
D[Step 4: Validate outputs]
E{Step 5: Results meet expectations?}
B --> C --> D --> E
E -- No: Debug & iterate --> B
E -- Yes: Release --> F[Release]
end
F --> Q
Q -- No: Done --> G[Complete]