Qbo Multicompany MCP

qbo-multicompany-mcp

A read-only MCP server that serves several QuickBooks Online company files to one AI session over stdio. Every call to QuickBooks names its company through an explicit entity key. There's no default company and no matching on company name.

Unofficial. Not affiliated with or endorsed by Intuit. Each user registers their own Intuit app and

answers Intuit's questions for themselves.

Who it's for

• You run or keep books for more than one QuickBooks company and want an assistant to read across them without mixing them up: close the month, compare companies, trace a transaction, cross-check vendor payments.
• You build tools for AI agents and want a worked example of an MCP server designed around how a model actually behaves.

Other QuickBooks MCP servers exist, some from Intuit, and some serve several companies or write to the books. Compare them against what you need. This one never writes.

AI client  -- stdio -->  qbo-multicompany-mcp  -- GET, checked before send -->  QuickBooks Online

| +-- entities.yaml your companies, by key +-- encrypted store one token chain per company

Quickstart on sandbox

You need Python 3.11+, uv, and an Intuit developer app with http://localhost:8765/callback registered as a development redirect URI.

git clone https://github.com/dkaleganov/qbo-multicompany-mcp.git

cd qbo-multicompany-mcp && uv sync --frozen && cp entities.example.yaml entities.yaml && cp .env.example .env && chmod 600 .env entities.yaml

Put your app's Client ID and Client Secret in .env, and your sandbox companies in entities.yaml. Then create the store key and connect each company in a browser:

uv run qbo-multicompany-mcp-auth --env-file .env --entities entities.yaml init-key

uv run qbo-multicompany-mcp-auth --env-file .env --entities entities.yaml connect acme_main

A first connect usually reports a name mismatch: answer y and copy the printed display_name into entities.yaml. Then add the server to your client with absolute paths. --scope user makes it available to every Claude Code session on your account, so you can run those sessions outside the repo folder:

claude mcp add --scope user qbo -- "<repo>/.venv/bin/qbo-multicompany-mcp" --env-file "<repo>/.env" --entities "<repo>/entities.yaml"

docs/getting-started.md walks through every step, including Claude Desktop and Codex. Production needs more: docs/production.md.

From PyPI, without a clone

uvx runs the published package. Save .env.example as .env and entities.example.yaml as entities.yaml in a private folder, chmod 600 both, and fill them in as above. With standing for that folder's absolute path:

uvx --from qbo-multicompany-mcp@0.1.1 qbo-multicompany-mcp-auth --env-file "<dir>/.env" --entities "<dir>/entities.yaml" init-key

uvx --from qbo-multicompany-mcp@0.1.1 qbo-multicompany-mcp-auth --env-file "<dir>/.env" --entities "<dir>/entities.yaml" connect acme_main

claude mcp add --scope user qbo -- uvx qbo-multicompany-mcp@0.1.1 --env-file "<dir>/.env" --entities "<dir>/entities.yaml"

Tools

Eighteen by default, and one opt-in. docs/tools.md has every parameter and a "Choosing a tool" table.

What makes it work well with AI agents

A model will call the wrong company, trust an empty result, and retry a bad value. The server is built around that.

• An explicit company on every call. With no default, a model has to name the company each time, so a wrong choice shows in the transcript. list_entities gives the keys, and a guard checks the company's name on first use in each process.
• Results name their company and age. Every answer from QuickBooks carries the entity, environment, the last four digits of the realm id (Intuit's Company ID), and when it was fetched, so a model can say which file a number came from.
• One list vocabulary. Every list that walks QuickBooks's pages returns count, rows, truncated, duplicates_dropped and pages_walked, so a model learns the shape once.
• Partial reports say so first. A problem in a cell or a row makes the result partial, and the text a model reads opens with PARTIAL RESULT. An empty report never establishes that there's no data, and list_transactions re-checks a filtered request with zero rows, unfiltered, before it reports "none".
• Refusals name the fix and never echo a value. A refusal names the parameter and the reason, usually before any request, and no text, id, date or number you sent comes back; it can name a configured entity key or one of this server's own option values. A test sends a marked value through every parameter of every tool and finds it repeated in no refusal.
• Company text is flagged as untrusted. Names, memos and report cells can carry anything. Every tool that returns them says in its description that they're untrusted input to the model and that the data can reach its provider.
• Sign-in stays with a person. Consent happens in a terminal and a browser, secrets come from the environment, and no tool takes a secret or an authorization code or returns one. Anything running as you can read a plain .env and the token store's key, from any folder. Keeping AI sessions out of the repo folder keeps .env and entities.yaml out of an agent's working folder; an agent with a shell can still reach both, and the key, as SECURITY.md explains.
• Descriptions route. Every tool's description says when another fits better, and the "Choosing a tool" table in docs/tools.md covers the whole surface for people.

Read-only, and how that's enforced

Intuit offers no read-only accounting scope: the one scope covers reads and writes alike. So read-only lives in this code. Every request is inspected as a fully built httpx.Request immediately before it's sent, and accounting requests may only be GETs to an allowlisted path on the exact bound company. OAuth calls go only to Intuit's fixed endpoints, the opt-in attachment download goes only to a host you pin, and everything else is denied. docs/design.md has the policy, and how walks, reports and the token store behave.

How it's verified

• Tests. uv sync --frozen --extra dev, then uv run pytest -q: 2,805 pass and 4 skip, with no credentials and no network. That count needs git 2.32 or later and gitleaks on your PATH; without gitleaks 25 more skip.
• Mutation gate. A passing suite doesn't prove a safety behaviour is protected, so scripts/mutation_gate.py has 480 entries that each delete one and require its tests to fail. It has controls of its own, so it can't quietly pass by testing nothing.
• Leak checks. Custom gitleaks rules for Intuit credentials, each proven to fire on a canary, plus scripts/scrub_check.py for identifier-shaped values.
• Live runs. Every tool has run over stdio in sandbox: the default tools on this release, and get_attachment on a pre-release build. Every default tool ran on production, read-only, on 2026-09-28 with a pre-release build, apart from TaxSummary under run_report, which has never returned rows. What the live runs found is in docs/api-findings.md.

CHANGELOG.md separates what was verified live from what has only run on fixtures.

Example prompts

Which companies can you see, and is each one connected?

Compare profit and loss for acme_main and acme_ops, this fiscal year-to-date against last. Say where each

number came from.

Show the balance sheet for acme_ops as of the end of last quarter, then the journal entries behind any

account that moved by more than 10%.

Which vendors in acme_ops carry the 1099 flag?

Give me candidate vendor payments for acme_main for 2025, and what needs review.

Known limits

• Read only, and QuickBooks Online only. There are no writes, and no QuickBooks Desktop.
• Transactions have their own tools. On transaction types run_query answers COUNT(*) and refuses SELECT *, and list_changes refuses them too. Use list_transactions and get_transaction.
• Filters validate differently. An unknown parameter name is discarded upstream, a bad enum value is a 400, and an id or text that matches nothing is an empty 200. This server handles each, and docs/design.md explains how.
• Vendor payment totals are candidates. They're a cross-check against a bank ledger, not a filing classifier, and they haven't been reconciled by hand against one.
• Sandbox isn't production. A value confirmed in sandbox may not hold on a production company.
• Some columns are unobserved. Class, department and currency report columns are documented, but no live response has shown them.
• `get_attachment` is only proven in sandbox. Production's storage host hasn't been observed, and the host check is a pre-check, not a binding to the address con