Zero Cost and MCP

Zero cost: your AI runs the engine

REGEN has no paid API keys and no hosted model. The intelligence comes from the AI client you already use (Claude Desktop / Claude Code, Cursor, any MCP-capable assistant), connected to the engine over the Model Context Protocol. The engine is the hands; your assistant is the head.

flowchart LR
    U[You] -- "start applying" --> AI[Your AI client<br/>Claude / Cursor / ...]
    AI -- MCP stdio --> MCP[python -m engine mcp]
    MCP --> ENG[engine: discover · gate · compose · apply · inbox]
    ENG --> LOG[(events.jsonl<br/>append-only)]
    AI -- reads your inbox with its own tools --> MAIL[(email)]
    AI -- submit_codes / outcomes --> MCP
    WEB[Control room :3000] --> API[REST API :8787] --> ENG

Loading

Why this design

Concern	Hosted-LLM design	REGEN
Cost	per-token bills, API keys to manage	$0: uses the assistant you already pay for or run locally
Secrets	keys in .env, rotation, leaks	no model keys exist to leak
Privacy	resume + answers sent to a third-party API	data stays on disk; the assistant sees only what a tool returns
Truth	model can invent experience	writing is template + Fact Bank only (see Tailoring); the assistant orchestrates, it doesn't author claims

What the assistant can call

Read tools are always on; write tools need REGEN_MCP_WRITE=1 on your machine.

Tool	Kind	What it does
engine_status	read	verified counts (proof on disk), waiting items, boards watched, outcomes
applications	read	applications by status, newest first
human_queue	read	only what a human must do (legal attestations, captcha walls)
outcomes	read	classified replies: confirmation, rejection, OA, interview, offer, scam
memory_query	read	knowledge-graph neighborhood of a company, fact id or lane
job_queue	read	discovered jobs waiting, filterable
pipeline_status	read	each applier's latest results and jobs waiting for an email code
pipeline_run	write	start discover → gate → compose → apply in the background
submit_codes	write	deliver email security codes the assistant read from the inbox

The same operations exist in the CLI (python -m engine ...) and the control room, so nothing depends on any one assistant.

Evidence

• engine/connectors/mcp_server.py: the tool definitions above.
• tests/test_api.py: API parity for the read surface.
• No model SDK appears in requirements.txt; grep -ri "api_key" engine/ finds none.