Muse Code vs OpenCode
Meta's cheapest-in-class agent, locked to one lab, against an MIT-licensed agent that will run whichever model you point it at.
Updated September 4, 2026 · Independent comparison — Muse Code, OpenCode are separate products.
Muse Code
Meta's local-first terminal coding agent
Muse Code is Meta's terminal coding agent, in beta from 2026-08-05 and generally available on 2026-09-01. It installs as a single static binary and is local-first — the agent, its subagents and the session logs stay on your machine, with only model inference hosted. Its standout capability is parallelism: big jobs fan out to child agents in isolated git worktrees at roughly core count minus two concurrency, so simultaneous work never collides. The trade is that it runs Muse Spark 1.2 and 1.3 and nothing else. Plans start at $5/month, with pay-as-you-go on the Meta Model API and a Contributor tier roughly 12x cheaper in exchange for training rights on your code.
OpenCode
MIT-licensed terminal-first agent
OpenCode is MIT-licensed and terminal-first, with a TUI plus extensions for VS Code, Cursor and Zed, and a headless mode for scripting. It reaches 75+ models through Models.dev, accepts GitHub Copilot and ChatGPT logins as well as raw API keys, and offers two first-party billing options that are entirely optional: Zen pay-as-you-go and a $10/month Go subscription whose key works in other agents too. Nothing about it is locked — you can read the source, fork it, and bring any key. What you give up is a co-trained model: OpenCode is a harness, and its quality tracks whichever model you feed it.
Bottom line
This is a lock-in comparison. OpenCode is MIT-licensed, reaches 75+ models through Models.dev, accepts Copilot and ChatGPT logins, and its two first-party billing options are entirely optional — there is no way for it to strand you. Muse Code is proprietary end to end, runs only Meta's Muse Spark models, and has no bring-your-own-key path to another lab. What Meta offers in exchange is a model co-trained with its own harness, worktree-isolated parallel subagents that are genuinely good on large repositories, an OS-enforced sandbox that fails closed, and a $5 entry price. The data term is the thing to read carefully: Meta's cheapest tokens come from the Contributor tier, and that discount is paid for with the right to train on your inputs and outputs.
Openness & lock-in
| Feature | Muse Code | OpenCode |
|---|---|---|
| Harness licence | Proprietary | MIT |
| Model weights | Closed | n/a — bring your own model |
| Bring your own key | Meta Model API key only | Any provider key, plus Copilot and ChatGPT logins |
| Models available | Muse Spark 1.2 and 1.3 | 75+ via Models.dev |
| Switch models mid-session | ✗ | ✓ |
| Fork or self-host the agent | ✗ | ✓ |
Pricing
| Feature | Muse Code | OpenCode |
|---|---|---|
| Free to run | ✗ | Yes — the agent is free, you pay only for tokens |
| Entry paid plan | $5/month (Everyday Usage) | Optional — OpenCode Go at $10/month |
| Higher-usage plans | $15/month (3x), $50/month (10x) | Zen pay-as-you-go, or any provider's plan |
| Pay-as-you-go option | Meta Model API: $1.25 in / $4.25 out / $0.15 cached per 1M | Zen, or the token pricing of whichever lab you point it at |
| First-party billing is optional | ✗ | ✓ |
| Discount for training rights | Contributor tier: ~$0.10 in / ~$0.20 out per 1M | Depends on the provider you choose |
Agent architecture
| Feature | Muse Code | OpenCode |
|---|---|---|
| Parallel work | Child agents in isolated git worktrees, concurrency roughly core count minus two | Not published |
| Where the agent runs | Fully local; only the model is hosted | Fully local; only the model is hosted |
| Model co-trained with the harness | Yes — Muse Spark is trained against Muse Code | ✗ |
| Session recovery | Append-only event log, replay-exact and restart-safe | Session history |
| Plan before editing | /plan, /grill and /goal commands | Plan and build modes |
| Project instructions | AGENTS.md, falling back to CLAUDE.md | AGENTS.md |
Surfaces & platforms
| Feature | Muse Code | OpenCode |
|---|---|---|
| Editor extensions | ✗ | VS Code, Cursor and Zed |
| Headless mode | Scriptable binary; TypeScript SDK in developer preview | Yes |
| MCP support | Enabled by default — stdio and streamable_http | ✓ |
| Multimodal input | Image and video, plus voice mode and web search | Depends on the model you pick |
| Platforms | macOS and Linux; Windows needs WSL2 | macOS, Linux and Windows |
Safety & data
| Feature | Muse Code | OpenCode |
|---|---|---|
| Approval controls | on-request default; rm -rf and sudo always prompt | Permission prompts |
| OS-level sandbox | Seatbelt on macOS, bubblewrap on Linux; fails closed | Not published |
| Who sees your code | Meta, unless you self-limit; Contributor tier trains on it | Whichever provider you point it at |
The Verdict
Choose Muse Code if...
- ✓You want a model co-trained with its harness rather than a generic one
- ✓You want worktree-isolated parallel agents on large repositories out of the box
- ✓You want an OS-enforced sandbox that fails closed
- ✓You want image and video input, voice mode and web search in the agent
- ✓$5/month for a frontier-class model is the deciding number
Choose OpenCode if...
- ✓You want MIT-licensed software with genuinely no lock-in
- ✓You want 75+ models and the freedom to bring any key
- ✓You already pay for Copilot or ChatGPT and want to reuse that login
- ✓You want editor extensions for VS Code, Cursor or Zed as well as a TUI
- ✓You want first-party billing you can ignore entirely
- ✓You need native Windows support without WSL2
Frequently Asked Questions
Is Muse Code open source?
No. The Muse Code harness is proprietary and Muse Spark's weights are closed. OpenCode is MIT-licensed, so you can read it, fork it and self-host it, and it holds no claim over the models you run through it.
Can I run Claude or GPT models through Muse Code?
No. Muse Code runs Muse Spark 1.2 and 1.3 only, with no bring-your-own-key path to another lab. OpenCode reaches 75+ models through Models.dev and lets you switch between them mid-session, which is the single largest functional difference between the two.
Which is cheaper?
It depends on what you count. OpenCode's agent is free and its billing is optional — you pay only for whatever model you point it at, or $10/month for the Go plan. Muse Code costs $5/month minimum but that includes the model. For a developer with no existing key, Muse Code is the cheaper complete package; for one already paying for Copilot or ChatGPT, OpenCode adds nothing to the bill.
Does Meta train on my code?
On the standard tier, no. The Contributor tier is where the roughly 12x discount lives — $0.10 in / $0.20 out per million tokens against $1.25 / $4.25 — and Meta trains on your inputs and outputs at that price. Check which tier your key bills on before pointing the agent at a private repository.
What does Muse Code do that OpenCode does not?
Three things worth naming: it creates one git worktree per child agent so parallel work cannot collide, it enforces an OS-level sandbox (Seatbelt on macOS, bubblewrap on Linux) that fails closed rather than degrading, and it keeps an append-only event log that makes sessions replay-exact and restart-safe. OpenCode publishes no equivalent for any of the three.
1DevToolOpen harness, cheap model, one workspace
OpenCode's argument is that the harness should be replaceable; Muse Code's is that a co-trained model beats a generic one. Both are right, which is why running them together works. 1DevTool runs 14 CLI agents including OpenCode, Claude Code, Codex, Gemini, Cursor and Cline in persistent terminals in one window, each on its own key or subscription, with automatic fallbacks when one is rate-limited. Around them sits a 23-engine database client, an HTTP client, browser profiles, Docker and Git worktrees. From $29 one-time, no subscription.