The problem
New clients cost an agency weeks before the first euro is spent.
A 30-question checklist half answered, account access that never arrives, a brief written from memory, work that starts on revenue numbers instead of profit. The people who could fix it are the agency's most expensive hours.
How it works
Eight stages. Five human gates. The agent does the paperwork.
Your client's AI talks to our agent
One MCP URL, one scoped token. Their assistant answers 32 onboarding questions one per turn; every answer is stored append-only with provenance. Client text is data, never an instruction.
From answers to assigned work
A BRD by section, one work item per answer, one spec per section, tickets in dependency order, and a working draft built from the client's own figures.
The agent cannot approve its own work
Its tool list has no approve, accept, review, release or send. A builder who tries to review their own ticket gets a 409. A number that traces to no client answer fails the pre-check.
Local-first by construction
Every server binds 127.0.0.1, the console loads nothing from the network, and the SDLC service refuses to start if the model URL isn't loopback. Rules first, so it runs with the model down.
Data room
Drop a client's sales export and the Analyst agent answers "which products sell more on promotion, and which promotions lose money", from the file, by region and customer segment, as a spec with tickets.
Autopilot, when the owner delegates
Seven stage agents, each in its own process with nothing but its stage's files, take a spec to released work. The Dev agent never reviews what it built; every run is signed.
The money moment
Seasonal decor needs 3.4× to break even. The agent puts that next to every draft.
| Category | Margin (client) | Margin-only floor (derived) | Client break-even |
|---|---|---|---|
| Tableware | 58 % | ≈1.72× | 2.1× |
| Textiles | 46 % | ≈2.17× | 2.6× |
| Lighting | 41 % | ≈2.44× | 2.9× |
| Seasonal decor | 33 % | ≈3.03× | 3.4× |
Floors are derived by the agent from the client's own answers and marked ≈. A draft that ignores them does not pass review.
Run it
Three commands, no model required.
git clone https://github.com/carljouffroy/aceops && cd aceops uv sync make sdlc-sample # console on http://127.0.0.1:8792 · agent MCP on :8791
78 tests, CI on every push. Full details, the GB10 runbook and the agent tool list are in the README.