mesa-mcp
Generates QuickBooks-compatible accounting exports for ledger periods, including a double-entry journal and per-transaction audit details.
Generates Xero-compatible accounting exports for ledger periods, including a double-entry journal and per-transaction audit details.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@mesa-mcp how much did my agents spend this month?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
tercet
The book of record for machine payments. Every agent purchase is recorded as it happens (request → quote → signed authorization), matched against on-chain settlement, and every difference gets a name — delivered-but-uncollected, paid-without-delivery, replay, chain-orphan. tercet never touches keys, balances or the payment path: it observes and produces evidence.
A tercet is a stanza of three lines that belong together — here, every purchase is one: the request, the authorization, the settlement. Dante chained his tercets so none could be pulled out without breaking the poem; this ledger chains its entries the same way.
The Python package is still named mesa (working name); it renames to tercet
with the first PyPI release. Project docs under docs/ are written in Portuguese —
they are the build diary and per-phase design docs; code, schemas and this README
are English.
The app — the ledger, live, in six screens
uv run mesa-app # -> http://127.0.0.1:840001 blotter (what your agents bought, with derived state and the event chain of each purchase) · 02 tca (waste, dedup across agents, cost per delivered result) · 03 risk (budget per task tree, human approvals, the payer passport) · 04 lab (point-in-time spend-policy backtest with honest confidence intervals) · 05 books (three-way reconciliation, sealed periods with two third-party timestamps, per-counterparty statement, BR tax view) · 06 operations (run the audited engines from the screen — demo, collectors, exports — testnet only).
The app is read-only by construction: its Postgres session opens with
default_transaction_read_only=on, so it cannot write to the ledger even if it
wanted to — and tests/test_app.py proves it. Real data always; testnet and
synthetic rows are labeled on the screen itself. Press any pill for its legend.
Related MCP server: ACA PoC MCP Tools Server
The MCP server — the ledger as tools for agents
claude mcp add tercet -- uv run mesa-mcp # then ask: "how much did my agents spend?"Seven READ tools (closed list, proven by test): status_do_livro, gasto,
compras (filterable), compra (one purchase + its event chain), vereditos,
passaportes, fiscal. Same guarantee as the app: the session is structurally
read-only — the MCP server has no write path; tests/test_mcp_livro.py proves it.
Transport: stdio, local.
The exports — universal accounting + OECD CARF view
uv run python scripts/fase13/contabil_build.py # -> contabil/<yyyy-mm>/ (universal CSV, QuickBooks, Xero + per-tx audit detail)
uv run python scripts/fase13/carf_build.py # -> fiscal/carf/<year>/ (OECD CARF view, born as OECD11 test data)A double-entry journal per period (debits == credits proven; micro-payments
aggregated honestly, with a 6-decimal audit bridge tying every cent to a tx hash),
and the CARF view — what a reporting crypto-asset service provider would report
about these transactions under the OECD framework (official July 2025 guide) —
labeled as a demo, synthetic identities, real numbers. Details and caveats:
docs/fase13-export.md.
From zero, with only this README
Prerequisites: uv, Docker, Python 3.11+.
# 1. dependencies + database (Postgres 17 on 127.0.0.1:5433 — your machine only)
uv sync
docker run -d --name mesa-pg -e POSTGRES_USER=mesa -e POSTGRES_PASSWORD=mesa `
-e POSTGRES_DB=mesa -p 127.0.0.1:5433:5432 `
-v mesa-pgdata:/var/lib/postgresql/data postgres:17
uv run python scripts/check_db.py # migrations + hash-chain genesis, idempotent
# 2. TEST wallets (written to a local .env; never committed) + faucet USDC
uv run python scripts/setup_wallets.py
# -> paste the buyer address into https://faucet.circle.com (USDC / Base Sepolia)
uv run python scripts/check_balance.py # balance read on-chain
# 3. a toy seller + 10 REAL x402 payments on testnet
uv run uvicorn mesa.http.seller:app --port 8402 # terminal 1
uv run python scripts/fase1/buyer.py 10 # terminal 2 (writes the ledger)
# 4. the chain is the source of truth: collect and reconcile
uv run python -m mesa.collector # matches (authorizer, nonce) with the ledger
uv run python -m mesa.cli # the verdict table, orphans in red
# 5. the product
uv run mesa-app # -> http://127.0.0.1:8400Costs zero real money: faucet USDC, and under the exact scheme the
facilitator pays gas.
What happened, what it proved, what comes next: docs/DIARIO.md (Portuguese) — the per-task build diary; per-phase design docs live in
docs/faseN.md.Secrets (the private keys
setup_wallets.pygenerates) go to a local.env(gitignored, never committed). To keep them elsewhere, pointMESA_ENV_FILEat the file — and keep it out of synced folders (OneDrive/Dropbox): a key in a synced folder is a key in the cloud.
Repo map
Path | What it is |
| Constants (USDC, network, chain id) + |
| Ledger access — insert-only by construction |
| The heart: three-way reconciliation, pure function |
| On-chain collector: |
|
|
| HTTP rail: toy seller (with chaos modes) + instrumented buyer |
| MCP: paid-tool rail (seller/buyer) + |
| The web app (FastAPI + Jinja2, server-rendered, read-only session) |
| The payer passport: portable, self-signed settlement history |
| The export renderers (accounting / OECD CARF / Brazil) |
| Numbered plain SQL — schema agnostic to rail and transport |
| Environment utilities + |
| The gates as pytest: dirty data in → expected classification out |
| Build diary + per-phase design docs (Portuguese) |
| Standalone verifiers — check the ledger and passports without trusting this repo |
Environment
uv sync
docker run -d --name mesa-pg -e POSTGRES_USER=mesa -e POSTGRES_PASSWORD=mesa `
-e POSTGRES_DB=mesa -p 127.0.0.1:5433:5432 `
-v mesa-pgdata:/var/lib/postgresql/data postgres:17
# 127.0.0.1 is mandatory: without it the ledger is open to any machine on the same
# Wi-Fi (docs/seguranca.md, hole #1). scripts/saude.py checks this, always.
uv run python scripts/setup_wallets.py # generates the EOAs -> .env (NEVER committed)
uv run python scripts/check_db.py # Postgres up + migrations + genesis
uv run python scripts/check_balance.py # USDC balance read on-chain
uv run python scripts/saude.py # ONE command: ruff+mypy+pytest+verifier+ports+secret scan
uv run python scripts/backup_db.py # ledger dump -> backups/ (routine, ~5s)Faucet: https://faucet.circle.com → USDC → Base Sepolia (20 USDC / 2h per address).
Anthropic key (only for the invoice-rail agent demo): ANTHROPIC_API_KEY=... in
your secrets env file.
Run the proofs of each phase
# Phase 1 — the ledger closes on dirty data
uv run uvicorn mesa.http.seller:app --port 8402 # terminal 1: HTTP seller
uv run python scripts/fase1/chaos_run.py # terminal 2: chaos + reconciliation + asserts
# Phase 2 — x402 over MCP
uv run python -m mesa.mcp.seller # terminal 1: MCP seller (port 8403)
uv run python scripts/fase2/mcp_once.py # terminal 2: 1 paid tool call + ledger
# Phase 10 — the payer passport (emits, verifies, gates a seller)
uv run python scripts/fase10/gate10_demo.py
# Any time: sweep the chain and reconcile
uv run python -m mesa.collector # testnet (seller payTo)
uv run python -m mesa.collector --pagador # mainnet (payer-side)
uv run python -m mesa.cliQuality ("done" has a definition)
uv run python scripts/saude.py # ONE command: ruff+mypy+pytest+verifier+ports+secret scanThreat model, and what is out of scope — by name: docs/seguranca.md.
Available Tools
7 toolscompraB
Uma compra específica (por rid, vindo de 'compras') com a cadeia de eventos dela — a gaveta do blotter.
Inputs: rid (required). Returns: object
comprasC
As compras do livro, mais recentes primeiro, com filtros: estado (um de liquidado, pago-sem-entrega, entregue-sem-cobrar, cobranca-pendente, expirou-sem-uso, autorizado-pendente, sem-pagamento, fatura-conciliada, fatura-pendente, ingerido), rede (CAIP-2), trilho (x402/invoice/mpp/ap2), agente, contem (texto no domínio) e limite (1-100, padrão 20).
Inputs: rede, agente, contem, estado, +2 more. Returns: object
fiscalB
A competência fiscal BR (DeCripto, F11): nº de saídas mainnet, total em R$ pela PTAX persistida e o veredito do limiar de obrigação PF. Sem ano/mes usa a competência atual (data de SP).
Inputs: ano, mes. Returns: object
gastoB
Resumo do gasto no livro inteiro: real (mainnet) vs testnet vs invoice, por trilho, por agente, por dia, e o desperdício (mesmo recurso, mesmo byte, comprado de novo).
Inputs: no named parameters. Returns: object
passaportesB
Os passaportes de pagador (F10) emitidos neste checkout, RE-VERIFICADOS offline nesta chamada: íntegro?, aceito pela política?, motivos nomeados, métricas assinadas.
Inputs: no named parameters. Returns: object
status_do_livroA
Até onde o livro está atualizado: cursores do coletor (bloco por rede) e a corrente de hash (seq/elos). Chame antes de confiar em qualquer número — o livro nunca finge tempo real.
Inputs: no named parameters. Returns: object
vereditosC
A reconciliação de três pontas (pedido × autorização × chain) feita AGORA: contagem por veredito nomeado, com a explicação de cada um. Uma diferença nunca é um gap silencioso.
Inputs: no named parameters. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 7 tools
Each tool targets a distinct aspect of the financial domain: status check, spending summary, purchase listing, purchase detail, reconciliation verdicts, payer verification, and tax calculation. No two tools have overlapping purposes, and even the singular/plural pair (compra/compras) is clearly differentiated.
All tool names are lowercase Portuguese nouns (e.g., gasto, compras, vereditos). The only deviation is status_do_livro, which uses underscores and is more descriptive. While the pattern is consistent in being noun-based, the mixed style (underscore vs. single-word) is a minor inconsistency.
With 7 tools, the server is well-scoped for a financial reporting and reconciliation service. Each tool covers a necessary function without being overly granular or bloaty, fitting comfortably in the optimal 3-15 tool range.
The tool set covers the full read-only lifecycle for the domain: status checking, spending summaries, purchase listing and detail retrieval, three-way reconciliation, passport verification, and tax computation. There are no obvious gaps for the server's apparent purpose of audit and analysis.
Related MCP Connectors
Read your Golem fleet's ledger from any MCP client: agents' sessions, live activity, signed record.
Read-only gateway for durable agent identity, consent, recognized work, and signed receipts.
Read-only discovery for exact-commit Agent Skill validation, x402 payment, and signed receipts.
Agent memory and a signed spending ledger: pay any x402 API, keep a provable receipt.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables agents to perform financial operations including categorization, reconciliation, and reporting on a tamper-evident, self-hosted ledger.AGPL 3.0
- AlicenseNot gradedqualityDmaintenanceProvides financial audit reporting and ledger query tools for AI agents, enabling generation of compliance, risk, and transaction reports and querying of transaction data, balances, and summaries from an in-memory SQLite database via MCP.MIT
- FlicenseNot gradedqualityCmaintenanceProvides a stateless Agent Chat Protocol bridge that lets agents inspect proposed payments against policy, retrieve public signed receipts, and verify their cryptographic proofs without wallet or payment capabilities.-
- AlicenseAqualityBmaintenanceEnables read-only, keyless queries of timestamped, definition-backed data about AI agent payments and activity on Base, including rankings, agent records, detections, metric histories, and evidence-pack pricing, from MCP clients or the terminal.25Apache 2.0