StanyMwamba/crewforeman-mcp

CrewForeman 3.11.0 is a local stdio MCP server. It runs Harness on your machine, sizes a team, builds in waves, has an independent reviewer score the result, and saves a record of every run.
## Roles
| Role | Who | What it does |
| Builder | DeepSeek Harness (default) | Sizes the team, builds in waves, revises on notes. |
| Independent reviewer | Configurable, any vendor (e.g. Kimi) | Scores 0–10 and lists blocking defects. |
| Client / coordinator | Your MCP client (e.g. Claude) | Sends the brief and reads the result. |
Using a reviewer from a different vendor means a model never grades its own work.
Use a reviewer from any vendor - see docs/REVIEWER.md.
The lead sizes the team and merges the best candidates - see docs/TEAMS.md.
Runs survive restarts and can be resumed - see docs/OPERATIONS.md.
## Method
CrewForeman creates one work item for each brief. If the same work is already running, it returns that run. Otherwise it sizes a team and plans waves, up to the number of helper agents Harness allows in parallel. A reviewer scores the result from 0 to 10 and lists defects that block acceptance. quality_rounds caps the revisions.
mermaid
flowchart LR
A[Request optional idempotency_key] --> B{Admit + work_id active duplicate?}
B -- yes --> R[Return existing run]
B -- no --> D[New attempt classify + bounded waves]
D --> E[Independent review]
E --> F{Pass?}
F -- no --> G[Revise quality_rounds] --> E
F -- yes --> H[Journal + result MCP task completes]
## Outcomes
mermaid
flowchart LR
A[Team builds] -- done --> R{Review score ≥ 8.5?}
A -- stalls --> X[Rescue team finishes the work]
X --> R
R -- yes --> OK([Accepted])
R -- no --> T{Tries left?}
T -- yes --> F[Fix with reviewer notes]
F --> R
T -- no --> N([Returned with notes])
classDef good fill:#e6f4ea,stroke:#1a7f37,color:#0f3d1e;
classDef warn fill:#fff4e5,stroke:#9a6700,color:#4d3300;
classDef step fill:#f6f8fa,stroke:#57606a,color:#1f2328;
classDef rescue fill:#ddf4ff,stroke:#0969da,color:#0a3069;
class OK good;
class N warn;
class A,R,T,F step;
class X rescue;
A run passes only at 8.5/10 or higher with no blocking defect. quality_rounds sets how many tries it gets. If the team stalls, a rescue team finishes the work. Time limits can stop a run. Accepted and returned outcomes update the journal and result.json. The MCP task completes only at a final state. See docs/STATUS.md for the status vocabulary.
Skills that learn: CrewForeman saves tasks that fit no existing skill. When a task recurs, it drafts a new checklist and has the reviewer check it against the 8.5/10 bar. A new skill is promoted only after five uses and at least three scored runs. Its median score must stay within 0.3 of the baseline. Its blocking-defect rate may rise by at most 0.05. It may lose at most one comparison against the baseline. auto_skills = auto | propose | off (see docs/skills.md).
## Convergence
Recorded runs, not a benchmark. One representative job (P3.2), round by round in its accepted run. Round 1 is a full review; round 2 re-checks what was flagged and the new changes.
| Round | What the reviewer checked | Review length | Score | Blocking defects |
| 1 | Full review | Long | 7.5 | 1 |