claw-forge CLI

claw-forge is a multi-provider autonomous coding agent harness.
Agents run in parallel, claim features from a dependency DAG, and keep working
until every feature passes — or you stop them.

GitHub: https://github.com/clawinfra/claw-forge
PyPI: https://pypi.org/project/claw-forge/

Install

CODEBLOCK0

---

Full Workflow

1 — Bootstrap a project

CODEBLOCK1

Creates:

• - CLAUDE.md — agent instructions (stack-aware)
• INLINECODE1 — slash commands (/create-spec, etc.)
• INLINECODE3 — provider pool config
• INLINECODE4 — API key template

2 — Write a spec

Create app_spec.txt (plain text) or app_spec.xml (structured XML):

CODEBLOCK2

Or use the Claude slash command inside Claude Code: INLINECODE7

3 — Plan: parse spec → feature DAG

CODEBLOCK3

• - Decomposes spec into atomic features
• Assigns dependency edges (A must pass before B starts)
• Writes task graph to INLINECODE8
• Uses Opus by default (planning is the most critical step)

CODEBLOCK4

4 — Run agents

CODEBLOCK5

Default: 5 parallel coding agents, reads claw-forge.yaml.

CODEBLOCK6

5 — Check progress

CODEBLOCK7

Shows: passing/failing/running/pending feature counts, recommended next action.

6 — Open the Kanban UI

CODEBLOCK8

Opens a browser-based Kanban board at http://localhost:8888.
Shows real-time agent activity, feature status, costs, and logs.

---

All Commands

Command	Purpose
INLINECODE11	Bootstrap project (CLAUDE.md, config, slash commands)
INLINECODE12

Parse spec → feature DAG in state DB |
| run | Run agents until all features pass |
| status | Show progress and recommended next action |
| ui | Launch Kanban UI at localhost:8888 |
| dev | Start API + Vite dev server (for UI development) |
| add <spec> | Add features to an existing project (brownfield) |
| pause | Drain mode: finish in-flight, start no new agents |
| resume | Resume after pause |
| fix <desc> | Fix a bug using RED→GREEN reproduce-first protocol |
| merge | Squash-merge a feature branch to target |
| input | Answer pending human-input questions interactively |
| pool-status | Show provider pool health and current routing |
| state | Start the AgentStateService REST + WebSocket API |

---

Config: claw-forge.yaml

CODEBLOCK9

---

Edit Modes

str_replace (default)

Standard exact-text matching. Works well with capable models (Sonnet, Opus).

hashline (recommended for weaker models)

Content-addressed line tagging. Each line gets a 3-char hash prefix. Model references lines by hash — immune to whitespace/indentation drift. Benchmark: 6.7% → 68.3% success rate improvement on Grok Code Fast.

CODEBLOCK10

---

Brownfield (add to existing project)

CODEBLOCK11

---

Bug Fix Protocol

CODEBLOCK12

---

Common Patterns

Quick greenfield project

CODEBLOCK13

Cost-optimised run

CODEBLOCK14

Multi-provider failover

CODEBLOCK15

YOLO overnight run

claw-forge run --yolo --concurrency 8
# Come back in the morning — all features should be passing

---

Tips

• - Planning quality = outcome quality. Don't skimp on the planning model. Use Opus for plan.
• More concurrency ≠ always better. Each agent needs context. 3–5 is usually the sweet spot.
• --edit-mode hashline dramatically helps weaker/cheaper models that struggle with exact text matching.
• claw-forge status tells you exactly what to do next — trust it.
• claw-forge ui is the best way to monitor long runs — live logs per agent, cost tracking, dependency graph.
• Provider API keys go in .env (never commit). .env.example is the template.