contextforge

Table of Contents

• What It Is
• Why It Helps
• Core Capabilities
• How It Works
• Installation
• Usage
• Automatic Routing
• Command Reference
• Tool Families
• Documentation
• License

What It Is

ContextForge is an open-source MCP server, CLI, and Claude Code marketplace plugin that helps Claude stay oriented in real repositories. Instead of treating each prompt like a fresh crawl, it builds and reuses a local index of repository structure, symbols, graph edges, session events, research output, and a separate durable memory layer for long-term facts, diaries, and wake-up capsules.

That lets Claude answer broad repository questions with compact receipts, trace changes through indexed graph relationships, keep large shell output out of chat, and operate on files inside the repo without constantly falling back to noisy manual exploration.

The source tree now lives in src/**/*.ts, the hooks/tests/plugin bootstrap are TypeScript too, and the repo now passes a real npm run typecheck instead of relying on a transpile-only migration. The runnable CLI, MCP server, hooks, and tests are emitted under dist/ during npm install / npm run build.

⬆ back to top

Why It Helps

• Reduces chat-window waste by returning compact repo digests instead of replaying broad manual crawls.
• Gives Claude better repo memory through a persistent local index with full text-body coverage and line counts.
• Improves targeting for search, symbol lookup, architecture questions, blast-radius analysis, and git change mapping.
• Routes natural-language prompts like “what touched this branch”, “what role does this file play”, “show me src/foo.ts”, or “continue where we left off” onto the right forge_* tool more reliably.
• Adds a low-context research lane with forge_batch and forge_lookup so logs, diffs, and command output stay indexed locally.
• Adds a Mempalace-style layered memory stack with wake-up capsules, scoped recall, wing/hall/room navigation, diary checkpoints, temporal facts, and provenance-weighted hybrid retrieval across sessions.
• Provides repo-native forge_read, forge_write, forge_edit, and forge_bash operations for smaller, more controlled tool output.
• Supports multi-repo workflows with a registry, repo groups, graph summaries, generated artifacts, and MCP resources.

⬆ back to top

Core Capabilities

Repository Understanding

• forge_start warms ContextForge and reports index progress.
• forge_scan, forge_understand, and forge_walk answer broad repo questions from inventory, audit, and indexed memory.
• Exhaustive walk mode can open every repository file locally, read text bodies, scan binaries, and return a compact audit-backed summary.

Graph and Architecture

• Derived graph areas, flows, and schema summaries.
• forge_scope, forge_impact, forge_changes, and forge_rename use indexed structure instead of plain text search alone.
• Generated artifacts for repository maps, contracts, and wiki-style summaries.

Low-Context Research

• forge_batch runs one or more repo-local shell commands and stores the full output in ContextForge’s research index.
• forge_lookup queries that stored output later without replaying raw logs into the conversation.

Durable Memory