Overview
Updated 2026-08-29
Frankie is the Software Factory — a specialized, long-running coding agent that runs on APEX. You hand it a pull request that carries a plan or spec, and Frankie drives that PR to green, review-clean, and self-verified: implemented, lint/build passing, CI green, reviewer comments addressed — then it stops and hands off to a human to merge.
Humans define the plan. Frankie implements the work, drives CI to green, and hands the result back for review.
What Frankie does
- Meets a PR wherever it is — implement from a plan, finish a partial PR, or act as a reviewer on a complete-but-red PR.
- Runs a durable 7-stage pipeline that survives pod restarts (state persisted to BigQuery, resumed via orphan-reclaim).
- Fixes actionable CI failures and code-review findings automatically, and re-pushes until it's ready.
- Never merges and never flips draft → ready-for-review — both are always a human action.
At a glance
| Built on | APEX runtime (dispatcher → BigQuery queue → worker pod) |
| Pipeline | 7 stages · durable resume |
| Watch TTL | 6h (configurable) · max 10 iterations |
| Merges | 0 — human only |
| Kill switch | FRANKIE_ENABLED |
The tool surface
Your coding agent stays the planner and conversational interface; the Frankie MCP handles explicit factory operations. The label is the API; the conversation is the interface. The current surface:
frankie_handover, frankie_finish, frankie_update, frankie_status,
frankie_thinking, frankie_takeover, frankie_cancel, frankie_preflight,
frankie_feedback, and frankie_help (the self-describing entrypoint).
Where to go next
- Getting started — connect the Frankie MCP and hand over your first PR.
- Architecture — the dispatcher, queue, worker, and 7-stage pipeline.
- FAQ — what to hand over, labels, troubleshooting.
Track all runs in the Apex Console, and
reach the team in #swat-factory.