Show HN: Forge – 3MB Rust binary that coordinates multi-AI coding agents via MCP

[vipdestiny]

I've been using Claude Code, Codex CLI, and Gemini CLI on the same projects. They're each great alone but running them concurrently is chaos: merge conflicts from simultaneous edits, decisions made in one context lost to another, and slow architectural drift as each agent optimizes locally without a shared plan.

Forge is an orchestration layer that sits between AI coding tools and your codebase. It's a single Rust binary (~3 MB, zero runtime deps) that runs as an MCP server over stdio. Any MCP-compatible AI tool can call it.

What it does: - File locking: When an agent claims a task, Forge locks the target files. Other agents see the lock and work on something else. Conflicts become structurally impossible. - Knowledge flywheel: Agents call forge_capture_knowledge to store decisions, patterns, and gotchas. Other agents query the store before making decisions. Knowledge compounds across sessions instead of evaporating. - Drift detection: Sends recent changes + the project spec to an LLM for alignment scoring. Catches "you were supposed to build auth but you're refactoring CSS" before it compounds. - Governance: 5-dimension health check (tests, security, docs, architecture, git hygiene) that agents and humans can query at any time.

The brain is pluggable: a free heuristic engine (pattern matching, works offline) or an LLM engine (GPT-4.1 by default). Switch with one CLI command.

State is a single JSON file in .forge/ — human-readable, git-trackable, zero operational overhead.

51 tests (30 unit, 9 CLI, 12 MCP protocol), 0 compiler warnings, 0 unsafe blocks.

MIT licensed. Whitepaper with the full architecture: https://nxtg.ai/insights/forge-whitepaper

Happy to answer questions about the Rust implementation, MCP protocol design, or the multi-agent coordination problem in general.

[YaraDori]

Really interesting approach to the coordination problem. File locking + knowledge flywheel is a clean pattern.

I've been thinking about the adjacent problem: even once you solve coordination, each agent still needs to know HOW to do the workflow. Right now that means someone writes detailed instructions or the agent fumbles through with generic prompts.

We've been experimenting with screen recording as a way to capture workflows (https://skillforge.expert) — you record yourself doing the task, and AI extracts every click/keystroke into a structured skill file. The agent replays it autonomously.

Curious: with Forge's knowledge flywheel, have you thought about storing reusable skill definitions that agents could query before attempting a task? Feels like the "capture once, replay many" pattern could stack well with multi-agent orchestration.