@invaro/opentax

The verifiable US tax oracle for AI agents — a Model Context Protocol (MCP) server where every answer is computed from a content-addressed corpus of cited tax rules, ships its assumptions, and is independently re-verifiable. Never let a model guess at tax law again.

🌐 opentax.invaro.ai · GitHub · AGPL-3.0 + commercial dual license

Why

On TaxCalcBench TY25, a Claude Sonnet agent using this oracle computes 48/50 complete returns exactly (96%, 98.2% of all lines) — versus 58% for the best web-search-equipped frontier model and 6% for the same model unaided. The two remaining cases contain reference-data inconsistencies the engine's line-level reconciliation surfaced (reported upstream).

Use the hosted connector (no install)

Paste this URL into Claude.ai (Settings → Connectors → Add custom connector), ChatGPT, or Cursor:

https://opentax.invaro.ai/mcp

Or run it locally (stdio)

claude mcp add opentax -- npx -y @invaro/opentax

Claude Desktop / Cursor config:

{
  "mcpServers": {
    "opentax": { "command": "npx", "args": ["-y", "@invaro/opentax"] }
  }
}

The same binary is also the full CLI and the self-hostable connector:

npx -y @invaro/opentax eval --status mfj --wages 120000 --kids 2   # terminal CLI
npx -y @invaro/opentax check --wages 120000 --expect 5746          # verify a claimed number
npx -y @invaro/opentax lookup standard deduction                   # citations behind an amount
npx -y @invaro/opentax mcp-http                                    # self-host /mcp on $PORT

One package: run with no arguments from an MCP client and it's the stdio server; give it a command and it's the CLI.

Tools

| Tool | What it does | |---|---| | calculate_tax | Any federal target (net tax, AGI, EITC, AMT, SE tax, …) from typed facts | | compute_return | The full Form 1040 line set from raw document transcriptions (W-2s, 1099-Rs, SSA-1099s) | | compute_state_return | IL-1040, VA 760, CA 540, NY IT-201 printed-form line sets | | verify_tax_claim | Gate a claimed number before presenting it: verified / refuted, with the difference | | find_tax_cliffs | Exact-cent benefit cliffs over an income range | | compare_filing_statuses | The same facts across all five filing statuses | | search_tax_rules | Full-text search over the encoded law: rules with citations, excerpts, and effective windows — coverage answers for research, zero hits means "outside the corpus" | | list_input_facts, explain_rule, lookup_tax_parameter | Discover inputs; read any rule's citation, formula, and validity window | | calculate_business_tax, calculate_fiduciary_tax, determine_dependent, verify_fact, is_tipped_occupation | Form 1120, Form 1041, § 152 dependency, proofs, § 224 tips |

What "verifiable" means

Every computation returns its assumptions (documented defaults it relied on), citations (statute source, section, URL on every rule), and a content hash of the exact rule corpus used (tax years 2025 and 2026, kept current — e.g. OBBBA amounts, state 2025 legislation). Answers are derivation trees you can re-execute, not model output.

The engine refuses rather than guesses: missing facts are named, uncovered situations return "not covered," and dates outside a rule's validity window are errors — never silent fallbacks to a stale year.