A multi-agent orchestration platform for AI coding agents featuring role-based coordination, Git Worktree isolation, automated merge review, and a unified web dashboard
Stoneforge is a multi-agent orchestration platform for AI coding agents, providing full lifecycle management through a Web Dashboard (default at localhost:3457) and CLI tool (sf). The project is at 0.1.0 early experimental stage, written in TypeScript with a pnpm + Turborepo monorepo architecture.
The platform features four built-in roles: Director decomposes user goals into prioritized, dependency-aware task plans; Ephemeral Workers are auto-spawned by the scheduling daemon to execute coding tasks in isolated Git Worktrees and terminate upon completion; Persistent Workers handle exploratory or one-off tasks; Stewards manage merge review, document scanning, stuck task recovery, and custom scheduled workflows.
The core architecture employs a dual-storage model — JSONL as the single source of truth (Git-tracked, append-only, diff-friendly, mergeable across branches) and SQLite as a caching layer (fast queries, FTS5 full-text search, materialized views), where SQLite can be rebuilt from JSONL at any time. Each Worker operates in an isolated Git Worktree at path agent/{worker-name}/{task-id}-{slug}, eliminating file conflicts at the source. The automated merge pipeline has Merge Steward run test commands — passing tests trigger squash-merge, while failures create a task handoff back to the queue with context.
The Web Dashboard provides a comprehensive control surface: Activity (real-time terminal output), Tasks (list and kanban views with six-stage state flow: Backlog → Unassigned → Assigned → In Progress → Awaiting Merge → Closed), Merge Requests (Agent PR review queue), Plans (task plan grouping), Workflows (reusable task sequences), Agents (registration, start/stop, topology graph), Workspaces (tmux-like terminal multiplexer), Messages (channel-based communication), Documents (versioned knowledge base), Metrics (throughput and efficiency indicators), and a built-in Monaco editor with LSP support.
Three Agent backends are supported — Claude Code (default), OpenAI Codex, and OpenCode — switchable via the --provider flag. Role prompts can be overridden at the project level through Markdown files in .stoneforge/prompts/. The workflow engine supports persistent state with failure-point recovery and configurable cron/event triggers. The platform aims to replace multiple standalone tools including Linear/GitHub Issues, Notion/Obsidian, Slack/Discord, manual Git workflows, and GitHub PR workflows for unified AI-assisted development management.
Installation & Quick Start:
npm install -g @stoneforge/smithy
cd your-project && sf init
sf serve
open http://localhost:3457
CLI Commands: sf init (initialize), sf serve (start server), sf daemon start (start scheduling daemon), sf agent register <name> --role <role> (register agent), sf agent start <id> (start agent), sf pool create (create concurrency pool).
Unconfirmed Information: No standalone website/documentation site found; actual published version of npm package @stoneforge/smithy not displayed on page; team/author info limited to org name "Stoneforge"; semantic search embedding model and vector DB details unspecified; no known production use cases.