Definition
What is MSP, the Muse Code Session Protocol?
The Muse Code Session Protocol, MSP, is the JSON-RPC interface a Muse Code host exposes over stdio when you run muse serve. It carries sessions, turns, tool calls, approvals and usage as typed events, so a client can render structure rather than parsing terminal text. Meta ships an MIT licensed SDK for it.
Last updated · Markdown version
What travels over it
- Sessions, keyed to a working directory.
- Turns, with the model and reasoning effort used.
- Tool calls, with the command and its output.
- Approval requests, with the mode that produced them.
- Usage, which is where cost and plan figures come from.
Why protocol beats scraping
A wrapper that parses TUI output breaks whenever the TUI changes and can never see anything the TUI does not print. A protocol client receives typed events, so approvals are objects you can queue, diffs are structured, and usage is a number rather than a guess.
Who can use it
Anyone. The SDK is MIT licensed, which makes client applications the intended path rather than a tolerated hack. That is the basis for saying Helicon is built the supported way even though it is unofficial.
Source: GitHub: meta-models/muse-code-sdk (SDK and protocol docs, MIT)
Frequently asked questions
Is MSP the same as the Agent Client Protocol?
No. ACP is a general protocol several editors use for any agent. MSP is Muse Code's own interface, so it carries Muse Code specific things like its exact approval modes and usage reporting.
Do I need to know about MSP to use a GUI?
No. It matters only because a protocol native client behaves better than one scraping a terminal.
Is the SDK open source?
Yes, MIT licensed.
Keep reading
- muse servemuse serve starts a Muse Code protocol host that applications can drive over stdio. What it is for, and how a GUI uses one host per workspace.
- the Agent Client ProtocolACP is a general protocol that lets editors talk to any conforming coding agent. How it differs from a Muse Code specific protocol, and what gets lost in the bridge.
- Zed with ACPRunning Muse Code inside Zed through an Agent Client Protocol bridge, compared with a standalone Muse Code desktop app. Editor native versus session native.
- an ADEAn ADE is the workspace you supervise coding agents from, the way an IDE is the workspace you write code in. What belongs in one and what does not.
- agent approvalsAn approval is the moment an agent asks before acting. The three Muse Code modes, what each one does, and why bypassing them is the wrong optimisation.
- a git worktreeA git worktree is a second working directory for the same repository. It is the cheapest way to stop two coding agents from editing the same files.
Same Muse Code. Same subscription. Better interface.
Free and MIT licensed. Windows installer, Apple Silicon, Intel and universal macOS DMGs, and a web build against a daemon you run.