Claude Code vs Muse Code vs Antigravity, compared

Comparison October 7, 2026

Tested 2026-10-07
Versions Claude Code 2.1.238 · Muse Code 1.4.3 (1.4.3-R5018.1) · Antigravity CLI 1.2.14
Scope surfaces and defaults only — none of the three was authenticated, so no task was run and nothing here is about output quality

Ask the three how many skills they ship with and you get twenty, zero, and a question about what you mean.

Muse Code     20 built-in skills, in the binary
Antigravity    0 bundled; empty plugin list on a fresh install
Claude Code    none in the binary; skills arrive as plugins from marketplaces

Three vendors, three different bets about where an agent’s capability should live. That one is visible in a single command, and so is most of what follows.

How these numbers were obtained

Every value below was read out of the tool rather than remembered. The useful trick is passing a deliberately invalid value and letting the parser print the valid set:

$ claude --effort bogus --print "x"
Warning: Unknown --effort value 'bogus' — ignoring it and using the default
effort. Valid values: low, medium, high, xhigh, max.

$ claude --permission-mode bogus --print "x"
error: option '--permission-mode <mode>' argument 'bogus' is invalid.
Allowed choices are acceptEdits, auto, bypassPermissions, manual, dontAsk, plan.

More importantly: every check was run against all three. That sounds too obvious to state, except that last week I wrote a post which explored one tool properly and compared it against what I assumed about another. Three claims in it were wrong, and all three failed the same way. The correction is at the bottom of that post.

So the discipline here is mechanical: no line in the table below exists unless the same command was run three times.

The table

Where capability lives

The twenty-versus-zero gap is the biggest single difference and it is not a measure of maturity.

Muse Code puts them in the binary. Twenty skills covering planning, git safety, Python environments, frontend taste, test collateral, scaffolding — and four for migrating off competitors. They are on by default. The bet is that most of what an agent should know is the vendor’s job to decide.

Claude Code ships none in the binary. Skills arrive as plugins, from marketplaces, with --plugin-dir and --plugin-url for local and remote ones and a plugin eval subcommand for running test cases against them. The bet is that capability is a distribution problem, and the vendor’s job is the registry and the tooling around it.

Antigravity ships nothing and says so plainly: agy plugin list on a fresh install returns “No imported plugins.” The bet appears to be that you bring your own, or that this is simply earlier.

Neither extreme is obviously right. Twenty bundled skills are twenty opinions you did not choose and cannot easily audit; an empty install is zero opinions and zero help.

The effort floor matters more than the ceiling

All three cap out around the same place. The spread is at the bottom:

Muse Code    none · minimal · low · medium · high · xhigh · max · ultra
Claude Code                  low · medium · high · xhigh · max
Antigravity                  low · medium · high · max

Only Muse Code offers none and minimal. That is a product expecting you to turn reasoning down for cheap work, not only up for hard work — which fits a pay-per-token model where the dial is a cost control rather than a quality control.

Permissions: two enums and a profile

Claude Code  --permission-mode       acceptEdits · auto · bypassPermissions ·
                                     manual · dontAsk · plan
Antigravity  --mode                  accept-edits · plan
Muse Code    --permission-profile    <ID>

Claude Code and Antigravity both model this as a per-session switch, one with six positions and one with two. Muse Code models it as a named profile you reference by ID — something written once and reused, and by implication distributed. muse config validating “enterprise configuration documents” points the same direction.

That is the clearest enterprise signal of the three surfaces.

The MCP difference is the one that will annoy you

All three support MCP. Only two let you manage it from the CLI.

$ claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
$ agy mcp add | remove | list | enable | disable
$ muse mcp login | logout          # OAuth only

Muse Code’s mcp subcommand does authentication and nothing else. Its own help says servers are “a streamable-HTTP entry under mcpServers in settings.json” — that is, you add servers by hand-editing JSON.

For a tool that bundles twenty skills so you do not have to configure anything, leaving MCP server registration as manual file editing is an odd gap. It is also the kind of thing that is invisible in a feature matrix and obvious within ten minutes of use.

Sandboxing

Muse Code    muse sandbox   — check or set up the OS sandbox
             --sandbox, --sandbox-network <MODE>
Antigravity  --sandbox      — run with terminal restrictions
Claude Code  no flag

Claude Code has no sandbox option. The word appears in its help only inside advice about other flags — “recommended only for sandboxes with no internet access” — which tells you the model is: the sandbox is yours to provide, and the tool’s own controls are the permission modes.

Six permission values instead of two or a profile ID starts to look like a deliberate trade rather than an accident.

Everyone imports from everyone, except the newest

All three ship a competitor importer. Laid out as a matrix, the pattern is clearer than any single row:

Two things fall out of it. Nobody imports from Muse Code, which launched in August and has not yet become something worth harvesting. And every vendor’s source list is a list of the incumbents it wants your users to be leaving — Antigravity imports from Gemini because it is the Gemini CLI’s successor, and from Claude Code because that is where the users are.

The commands are not symmetric in depth. Claude Code’s walks config and MCP servers and writes a reference skill for anything it cannot map. Muse Code’s spans five entry points including unfinished session state. Antigravity’s agy plugin import claude|gemini handles plugins, and on my machine reported “No claude extensions found” despite an installed-plugins directory being present — so either it looks elsewhere or it wants a format I do not have.

One each that the others don’t have

Muse Code — its own wire protocol. muse schema exports an MSP schema as JSON Schema or TypeScript; muse serve runs an MSP session host over stdio; muse session-message sends messages between sessions. Multi-agent coordination as a published contract, not an emergent pattern.

Antigravity — projects you select. --project <id|name> and --new-project make a project a session attribute you choose at launch. Claude Code also has projects, but they mean something else: claude project purge deletes “all Claude Code state for a project (transcripts, tasks, file history, config entry)”, so a project there is a state container keyed to a directory, not something you pick per session. Muse Code has neither. It also ships agy mic-serve, which serves your machine’s microphone to a CLI on another host, which I cannot place.

Claude Code — eval. claude plugin eval runs test cases against a plugin’s skills. It is the only one of the three that ships a way to check whether the thing you installed actually does what it claims.

What this does not tell you

Everything that decides which one you should use.

No task was run on any of the three. Nothing here speaks to code quality, instruction following, long-context behaviour, recovery from error, or cost in practice. A surface comparison measures what a vendor decided to expose, which correlates with the product’s intent and not necessarily with its competence.

Three snapshots on one day, too. Muse Code left beta five weeks ago; Antigravity’s CLI is newer than that. Checking these numbers against the versions at the top is the first thing to do before trusting any of them.

Corrections, 2026-10-07. Two rows were wrong, both the same way.

1. Antigravity has a competitor importer — agy plugin import claude|gemini. I had read agy --help and not its sub-subcommands. The who-imports-from-whom matrix above was added afterwards, because the full picture turned out to be more interesting than the row.
2. Claude Code has projects too — claude project purge manages per-project state. They are not the same concept as Antigravity’s, which is a session selector, but “neither of the others has it” was wrong.

After the second one I stopped patching and re-ran every remaining row against all three tools, including sub-subcommands. Worktrees, sandboxing, MCP management, permission models, effort tiers and the wire-protocol row all survived that check unchanged.

The pattern across all of today’s errors, here and in the previous post: I keep verifying the tool I am curious about and inferring the others. A comparison table needs every cell run, not every row reasoned about.