Concepts
How Bifra turns a source into a typed API.
Bifra is an agent that genuinely does the work — it studies a source, shapes an API, builds and deploys it, and verifies it works — by calling tools. Each phase is a typed contract, enforced inside the tool the agent calls.
The phases
Study
The agent drives a real browser (via Steel) and captures the data endpoints a
source calls, scoring and replay-probing them. When a site is server-rendered
with no data API, the agent probes further with bash (curl, sitemaps, CSV/JSON
endpoints) before deciding anything.
Spec
The agent designs the agent-facing API — tools shaped for how an agent wants to use them, not 1:1 mirrors of raw endpoints — and commits it. You approve the shape once.
Build
The approved spec is bundled into a Cloudflare Worker that serves a typed REST
surface, an MCP server at /mcp, and an OpenAPI document at /openapi.json.
Deploy
The worker is shipped live and its URL returned.
Verify
Structural checks run against the live API. If they fail, the agent reads the real error, adjusts the spec, and rebuilds — healing on its own until it works or reports a clear diagnosis.
Where the loop runs
The agent is the same everywhere; only its sandbox differs:
- Self-hosted — the Claude Agent SDK runs the loop in a sandbox you control.
- Managed — the same agent runs on Anthropic's Managed Agents sandbox.
One prompt, one toolset, one model. Provider secrets (Steel, Cloudflare) run in tool handlers on your side and never enter the model's context.
Every API is also an MCP server
Because rails serves /mcp alongside the REST routes, every API Bifra builds is
immediately usable by an agent over MCP — no extra step.