Canopy
Canopy (harness-canopy) is a modern, self-contained MCP (Model Context
Protocol) server and terminal UI for orchestrating AI agent sessions,
background tasks and file event triggers. One Rust binary, zero runtime
dependencies.
It turns a single machine into a small agent operations center:
- Schedule agents on cron expressions or file-system events.
- Run interactive agents in real PTYs with full terminal emulation.
- Coordinate multiple agents working in the same workspace.
- Remember — a project-scoped knowledge graph that persists facts, patterns and session summaries across sessions.
- Search your own documents with a local-first RAG pipeline.
- Automate multi-step processes with a DAG loop engine.
Everything persists in an embedded SQLite database under ~/.canopy/ —
no external services, no cloud account.
The pieces
| Piece | What it does |
|---|---|
| Daemon | MCP server (Streamable HTTP + stdio), scheduler, watcher engine, database |
| Canopy Hub (TUI) | Full-screen terminal UI for agents, loops and system metrics |
| 83 MCP tools | Agent management, sync, intelligence, seeds, loops, specs, queues, blueprints, RAG, projects |
| Seed identities | Persistent, evolvable agent personalities stored as TOML |
| Gamification | 28 missions across 6 categories tracking usage milestones |
How the documentation is organized
- Installation — install and set up the daemon.
- Quick Start — daemon, setup wizard, first agent.
- The TUI — Canopy Hub — the interactive terminal interface.
- Agents — interactive, background and terminal agents; seed identities.
- Intelligence & Sync — knowledge graph and multi-agent coordination.
- Loops — the DAG loop engine.
- Usage Patterns — the shapes a graph can take, and where each breaks.
- Recipes — those shapes built end to end, with real calls and prompts.
- RAG Pipeline — personal document search.
- MCP Tools — all 83 tools by category.
- CLI Reference — every
canopycommand.
Part of UniverLab
Canopy is an experiment of UniverLab, an open computational laboratory. It follows the lab’s engineering principles: one tool one job, reproducibility first, offline-friendly design.