How to Use Muse Code: Install, Sign In and Run a Task

Install and identify your Muse Code version

On macOS or Linux, use the installer linked in the official Muse Code quickstart. Then check the installed version before copying commands from a tutorial. The examples here were checked against the Muse 1.3.0 command surface.

curl -fsSL https://dev.meta.ai/install.sh | sh
muse --version
muse --help

Use a normal terminal in your own account. If the command is missing after installation, open a fresh terminal so it receives the updated PATH. Record the version when reporting a failure. For native PowerShell commands and platform differences, follow Muse Code on Windows.

Open a project and sign in

cd /path/to/your/project
muse --model muse-spark-1.3 --reasoning-effort medium

Replace the sample path with your checkout. Choose the browser sign-in option offered by Muse, or use your own Meta API key. The interactive /login command reopens authentication. An existing API key can take precedence over browser credentials; check Meta’s authentication guidance if the account or billing route looks wrong.

The explicit model flag makes this example reproducible. Some Muse setup pages still name Spark 1.2 as the default, while the current model catalog lists 1.3. Do not infer the active model from the age of an article. Read the model shown in your session and use the models, pricing and privacy guide before choosing a different tier.

Make the first request easy to verify

Read README.md and the project’s build configuration.
Explain how this application starts and identify one test command
that already exists. Cite the files you used. Do not edit files.

Choose a repository whose purpose you know. If it has no README, name the actual entry point or configuration file instead. Compare the response with the files on disk. A plausible test command is insufficient: it must exist in this checkout and apply to the package you intend to change.

For a second request, pick one observable defect, such as an incorrect validation message. Describe the input that triggers it and the expected result. Ask Muse to inspect its callers, make the smallest correction and run the relevant check. Review git diff yourself before accepting the result. This gives you a useful baseline for later model comparisons.

Add project instructions without replacing existing rules

muse init --dry-run

Preview the proposed instructions first. The dry run does not create a file. If there is no AGENTS.md and the proposal is correct, run muse init, then review the generated file. If the repository already contains an AGENTS.md, read and edit that file deliberately. Keep commands that help the agent work: the package directory, the test command and a restriction that matters. Avoid filling it with generic programming advice. Configuration and project context explains how Muse loads those files.

Trust the correct workspace when Muse asks. Project instructions and hooks are part of that trust decision. A prompt saying “do not edit” communicates your task scope, but it does not itself change the tool permission profile. Review Muse permissions when you need enforced read-only access.

Resume the original task and inspect its state

muse resume --last

Run this from the intended project when you return. Before asking for more changes, tell Muse what changed outside the conversation: a dependency update, another developer’s commit or a manual edit. Session history cannot guarantee that the current files match the earlier discussion.

If you interrupt an edit, inspect the diff before issuing a replacement request. A cancelled answer does not mean every file is untouched. Ask the agent to explain the current state, then continue from that evidence. The interactive controls reference covers session management and interruption.

When the first task works

Keep the setup small until you have a verified edit. Add MCP tools, skills or workflows when a real task needs them. For a repeatable review, save the exact prompt and acceptance condition; that is enough to compare results across sessions.