forge

Forge is a structured delivery loop tool for Claude Code, designed to transform a single-agent workflow that easily spirals out of control on complex tasks into an engineered process that is plannable, parallelizable, verifiable, retryable, and memory-capable.

Purpose

Forge targets scenarios where ordinary prompts are no longer enough. When a task involves multiple files, requires coordination between modules, or needs proof that it actually works, Forge provides Claude Code with a workflow, not just another prompt. It breaks work into modules, runs the parallelizable parts in parallel, strictly verifies each module, retries on failure with debugging context, and remembers what will be useful for the next task.

Key capabilities

• Structured planning: Break features into modules with dependencies and verification commands.
• Parallel execution: Run mutually independent modules simultaneously in isolated worktrees.
• Deep verification: Check files, commands, syntax, and cross-module API contracts.
• Smart retries: Track attempt counts, detect stagnation, and escalate to a debugging agent.
• Resumable sessions: Continue unfinished work instead of starting over.
• Cross-session memory: Remember conventions, failure patterns, and test commands.
• Status visibility: Show progress in Claude output and the terminal status line.

Forge uses DAG plans rather than linear chains, with worker agents running in parallel when dependencies allow. Through worktree isolation and automatic WIP commits between layers, it ensures each layer is written to disk before the next layer branches. Each module's verification commands and acceptance criteria are defined during the planning phase, preventing worker agents from quietly lowering the bar. Multi-perspective review consists of an independent review agent plus a 3x self-consistency check, used to catch cross-module defects that a single-module reviewer easily misses. A structured failure ledger records failure patterns in JSONL format and recalls them by pattern ID in the next plan.

Usage

After installation, restart Claude Code (if it is already open), open any repository you want to work in, and run /forge <objective>. The first launch may take a few seconds because Forge automatically bootstraps its MCP server dependencies.

The commands available to users include:

• /forge: the complete orchestrator workflow.
• /forge-status: the current plan, module progress, and learned patterns.
• /forge-validate: re-run verification for a module.

Forge ships with four focused agents:

• planner: explores the codebase and proposes a module plan.
• worker: implements one module in an isolated worktree.
• reviewer: checks correctness, security, and contract mismatches.
• debugger: investigates failed modules and drives retries.

The bundled MCP server provides shared runtime capabilities, including tools such as iteration_state, so multiple agents can coordinate without relying on loose conversational memory.

When to use it

Forge is suitable for scenarios where a task involves multiple files, requires coordination between modules, or needs proof that it actually works. It does not try to replace normal use of Claude Code; for many tasks, using Claude Code directly is enough. Forge is aimed at tasks where orchestration is essential.

Limitations and caveats

• The first launch may take a few seconds because it automatically installs MCP server dependencies.
• If Claude Code is already open, it needs to be restarted.
• Nothing is executed until the user explicitly approves the plan.
• Forge creates a local .forge/ directory in the working project, used to store plans, logs, memory, retry history, and resumable state.
• For manual installation, you need to use .claude/settings.json Connect the MCP server using the reference configuration in .
• The project is under the MIT license.

Forge's architecture includes the orchestrator, agents, MCP server, and state flow. The documentation also provides an MCP tool reference (including the input/output schemas and examples for all 7 tools), a changelog, contribution guidelines, and a security policy.