Setting Up Muse Code and Comparing Coding Plans

Muse Code is Meta's coding CLI, and today I got it installed and connected to Paseo before deciding where it fits beside the coding plans I already run.¹ This is a setup and research pass. I have not compared coding quality against anything yet, and nothing here is a benchmark.

The install is ordinary. The credential handling is where the surprises were. The first browser login reached Meta but the credential store failed inside the execution sandbox, and repeating the login with Keychain access succeeded. With no API key in the environment, the sessions use the account login, a live model check came back fine, and Paseo's native Muse provider detected the CLI and listed its models.²

Three things to settle before trusting it with real code

Contributor mode defaults on. It trades higher allowance for permission to train on prompts, code, files and responses, and the setting is not available on some account types.³ My smoke test ran with it off, but the account default is still on. That is worth checking before any private repository work, and it is a good example of a privacy default sitting behind a usage increase.

Subscription credentials do not travel. The subscription covers the Muse Code CLI signed in to the vendor account; additional API keys are billed separately pay-as-you-go, and a key in the environment takes precedence over the browser login.⁴ So the plan funds interactive work in its own harness. Calls from other tools are a different bill.

The quota meter reports nothing. The usage endpoint returned an empty result, so remaining allowance is unknown from the API. Local session logs are the only usage signal available.

What the terms comparison changed

I checked the published terms of the plans in my stack rather than assuming the pricing pages describe the same product. The billing units differ, but the use restrictions are the sharper distinction. One coding-tool token plan documents its quota as valid only inside supported programming tools, treating scripted or backend API use as a violation.⁵ Consumer Claude OAuth is documented for individual use, with third-party agent routing blocked since April 2026. ChatGPT plan usage is tolerated in preferred coding tools while programmatic workflows are directed to API keys. The practical rule I took away: interactive coding, including agent-harness work, is what these subscriptions are for, and unattended work belongs on pay-as-you-go keys.

The usage shape from my first local session logs supports that. In under two hours the harness made a little over 300 model completions across dozens of subagent calls, and about 97 per cent of input tokens were cache reads. These are local log totals, not a quota meter reading, but the shape is the useful part: a plan priced in prompts per few hours would be exhausted almost immediately at this workload, and a multiplier-style allowance is the only fit.

Open question

Whether Muse earns a permanent place beside the existing seats is unmeasured. The test I want is three representative tasks: a small fix on the fast model, a multi-file change on the main model, and one hard diagnosis on the top-tier model that only the premium plans include. That last one is the only way to tell whether the premium tiers earn their keep, and it has not run yet. For now this is an additional lane for interactive work, not a replacement for anything.