AI Tool Atlas
Plugin Official

Math Proof

Two skills for research-level math: solo (single session) and siege (hours of judge and worker sub-agents), ending in proof.md.

Author
Anthropic
Works with
Claude Code
License
Apache-2.0
Documented
Oct 2, 2026
Repository ↗

What Math Proof is

/math-proof:solo works a problem in your session with a notes file. /math-proof:siege runs rounds of judge and worker sub-agents for hours and states what is and is not proved. Both disable web tools and write proof.md.

What it does

How Math Proof plugs in: which agents run it and what it addsRUNS INYour agentClaude Codeone host documentedPLUGIN · OFFICIALMath ProofBundles of skills, commands, hook…4 install steps3 permission noteslicense: Apache-2.0WHAT IT ADDSWhat it doessolo: one long turn writing plan…siege: judge writes questions, wo…Two workers check a complete proo…proof.md includes a Status sectionagent → extension → new capability · from the official repository or docs
Where Math Proof runs and what it adds, from its documentation
  • solo: one long turn writing plan and results to a notes file
  • siege: judge writes questions, workers answer in parallel, judge keeps a ledger
  • Two workers check a complete proof line by line before conclusion
  • proof.md includes a Status section
  • Resumable by re-giving the same command in the same directory

When to use it

Good fit

  • Hard mathematics where solo is tried first
  • Long unattended proof attempts with a spend cap

Not for

  • Trusting literature citations; they are quoted from memory

How to install Math Proof

You need: Claude Code 2.1.280+; Python 3.7+ as python3 or python for siege

  1. Run Claude Code
  2. Run /plugin install math-proof@claude-plugins-official
  3. Add the env block to ~/.claude/settings.json
  4. In an empty directory run /math-proof:solo or /math-proof:siege
Configuration example (from the official docs)
{
  "env": {
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "128000",
    "CLAUDE_CODE_DISABLE_BACKGROUND_TASKS": "1",
    "CLAUDE_ASYNC_AGENT_STALL_TIMEOUT_MS": "14400000"
  }
}

Permissions and security

What it can access

  • Writes files under ./math-proof-solo/ or ./math-proof-run/
  • Judge runs Python via the shell tool
  • Switches web tools off during its run

Security notes from the maintainers

  • Use --dangerously-skip-permissions only in a disposable container or VM, non-root

Review the source before granting an agent access to credentials or production systems. We document what the maintainers state; we do not audit code.

Limitations

  • Siege can run well over a hundred sub-agents; check /usage
  • One problem and one run at a time per directory

Tools it extends

Frequently asked questions

What is the difference between math-proof solo and siege?

solo works in your session; siege runs rounds of judge and worker sub-agents for hours. source

How do I cap spend for math-proof siege?

With an API key, add --max-budget-usd to the unattended claude -p command. source

What env settings does math-proof need?

CLAUDE_CODE_MAX_OUTPUT_TOKENS, CLAUDE_CODE_DISABLE_BACKGROUND_TASKS and CLAUDE_ASYNC_AGENT_STALL_TIMEOUT_MS. source

Related guides

Sources and gaps

Not stated by the maintainers: Release date and version history are not shown in the plugin README.

More extensions to look at