OPEN-SOURCE AGENT RUNTIME · v0.1.13

Build the agent you actually want.

One open runtime for models, tools, Skills, and memory. Swap any layer without rewriting the rest — in your terminal, on your desktop, on any machine you trust.

Download Magi ↓Explore on GitHub ↗

Open source Local-first No vendor lock-in

magi · local workspace● Ready

Provider fallbackmagi / main

anthropic:claude-sonnet-4-6⌄

Make provider fallback observable, then add coverage for recovery.

✦

I’ll trace the router, expose fallback events, and verify behavior with targeted tests.

⌁

Inspecting provider routerRead 4 files · 1.2s

✓

›_

Running targeted testsvitest provider-fallback

Message Magi

＋变更前询问/magi

91 tools, on demand

✓ approval required

One runtime. Five replaceable layers.

ModelToolSkillHarnessMemory

WHY MAGI

Not another fixed agent.

Most agents ship with the model, tools, memory, and execution rules already decided. Magi opens those layers up — swap what doesn't fit and keep the runtime underneath.

01Model

Route any model

OpenAI, Anthropic, DeepSeek, any compatible endpoint. Aliases, auto-routing, and fallback for when a provider goes down.

02Tool

Load tools on demand

91 built-in tools — files, shell, Git, web, memory, sessions, sub-agents, multi-machine — loaded only when a task needs them.

03Skill

Package the workflow

Turn repeatable work into Skills: the steps, the tools, how to verify, and what to do when something fails.

04Memory

Keep knowledge reviewed

Sessions persist. New lessons land in LearningDrafts first — nothing rewrites long-term memory behind your back.

COMPOSABLE BY DESIGN

Change the layer.
Keep the runtime.

A research agent and a coding agent can behave nothing alike — and still share sessions, approvals, orchestration, and recovery underneath.

Model

Route and fall back

Tool

Act on the world

Skill

Reuse a workflow

Harness

Define behavior

Memory

Learn with review

agent.yaml

model: main
tools: [files, shell, git]
skills: [verify, review-pr]
harness: coding-agent
memory: project-reviewed

MAGI RUNTIME

The hard parts are already built.

Agent loopStreamingApprovalsRecoverySessionsSub-agentsRoutingDispatch

ONE CORE, MANY SURFACES

Use Magi where the work lives.

A daily coding tool, a desktop workspace, or the engine behind your own product. Same runtime, your call.

A focused workbench

See the work, not just the answer.

Projects, streamed activity, model switching, approvals, file changes, and terminal output — one local window, no tab juggling.

Learn more ↗

Refine provider fallbackmagi · main

● Local agent ready

Make provider fallback observable and add tests.

✦

I’ll trace the current routing path, update the events, and verify recovery.

✓ Read provider routing

✓ Patch fallback events

↻ Run targeted tests

Message Magi main · claude-sonnet-4-6 ↑

GET MAGI

Download and start building.

Desktop for the daily driver, CLI for the terminal. Same runtime either way.

Magi Desktop

v0.1.14

The full workbench — projects, task history, approvals, and a local agent service that starts itself.

macOS (Apple Silicon)165 MB · .dmg↓macOS (Intel)168 MB · .dmg↓⊞Windows (x64)126 MB · .exe↓

All downloads & SHA-256 checksums ↗

Magi CLI

v0.1.13

One command to install, runs in any repository. Needs Node.js 22+.

terminal

$ npm install -g @edwardlee5423/magi@next
$ magi

View package on npm ↗

CONTROL STAYS WITH YOU

Powerful enough to act.
Careful enough to ask.

Plans, commands, edits, and remote work go through one permission model with a full approval trail. You decide where it has to stop and ask.

Plan before executionBig changes show you their scope before they start.
Approve at the boundaryAllow once, make it a rule, or just say no.
Carry trust across devicesCheck status and approve from a paired phone.

⌁

Approval requiredShell · sensitive action

Magi wants to run a command outside the current project.

npm publish --access public

Working directory/Users/edward/code/magi

✓Nothing runs until you say so.

BUILD BEYOND CODE

One core.
Many agents.

Different work needs different tools, instructions, permissions, and definitions of done — not another runtime built from scratch.

Code

Understand repositories, edit across files, run commands, debug, and verify.

Research

Web, local documents, structured workflows, and context that lasts.

Teaching

Turn repeatable pedagogy into Skills with the right tools and guardrails.

Data

Domain tools, completion rules, and model routing — all configured per agent.

Operations

Hand off long tasks, watch progress, approve the sensitive parts remotely.

Your agent

Replace the upper layers. Keep the runtime underneath.

OPEN, LOCAL, COMPOSABLE

Start with Magi.
Shape what it becomes.

Use it as a coding agent today. Rebuild it into your own agent when you're ready.

npm install -g @edwardlee5423/magi@next⌘

Download Magi ↓Read the docs