Drivetrain MCP server: what it reads, what it drafts, and how to connect Claude or ChatGPT
Drivetrain hosts a per-tenant MCP server with 22 tools: 18 read metrics and data, 4 save boards or transformations as drafts. OAuth or API key. Verified 21 Oct 2026.
- Official MCP server · read & draft-only write, hosted per tenant
- Hosting
- Remote, hosted on AWS in US East (N. Virginia) and EU (Frankfurt), per Drivetrain. One endpoint per tenant: https://{tenant}.drivetrain.ai/drive/api/v1/public/atlas/sse.
- Auth
- OAuth with each user's Drivetrain login, or a Drivetrain API key sent as a header via mcp-remote (Claude Desktop). OAuth access tokens last 24 hours, per Drivetrain; the API key's expiry isn't documented.
- Read / write
- 18 read tools; 4 write tools that create or update boards and dataset transformations as drafts a person saves.
- Works with
- ChatGPT (developer-mode app), Claude.ai (custom connector), Claude Desktop (API key). Other clients not documented.
- Plan requirement
- No MCP price published. Per Drivetrain, the Claude OAuth route needs Claude Team or Enterprise; the API-key route works on free Claude.
- Last changed
- 2026-10-09 (MCP overview and tool reference updated: 22 tools, 4 draft-only write tools)
- Verification
- Verified 21 Oct 2026 by theFinanceStack · confidence: medium
What can an agent do with Drivetrain through MCP?
An agent can search metrics and pull their values by period, dimension, Version and Scenario, read Lists, boards, datasets and model structure, and draft boards or dataset transformations for a person to save.
Where it runs
| Type | Tools | How they behave |
|---|---|---|
| Read | 18 | Drivetrain says they never change your account. |
| Write | 4: create_board, update_board, add_dataset_transformations, update_dataset_transformations | Per Drivetrain's tool reference, none posts directly. Each saves a draft that takes effect only when you click Save in the editor or save the transformation draft. |
Good to know
- The same page's intro says "most" write tools save a draft, so treat every write as needing review.
- No tool moves money or approves spend.
- Drivetrain's product page (not reachable on 21 Oct 2026) has described Claude editing models and updating assumptions. The current tool reference lists no tool for either.
Can
- Search metrics and read their definitions and dimensions
- Fetch metric values by date range, dimension, Version (e.g. Actual, Budget) and Scenario
- Read Lists, existing boards and sections of the model structure
- Read connected datasets, columns and current transformation steps
- Check a planning formula's syntax
- Draft new or changed boards and dataset transformation steps (write, saved as drafts)
Cannot
- Return data the signed-in user can't see under their Drivetrain role (per Drivetrain)
- Apply a board or transformation change until a person saves the draft (per Drivetrain)
- Edit metrics, formulas or assumptions: no such tool is listed
- Export tool-call logs to a SIEM; Drivetrain says this is planned
What access does it get?
The connection acts as the Drivetrain user who signs in, with exactly that user's permissions; Drivetrain says it checks them on every tool call.
| Where | Scopes and permissions (as documented) | Admin requirement |
|---|---|---|
| Drivetrain MCP server, OAuth (ChatGPT, Claude.ai) | No Drivetrain-specific scopes. The OAuth metadata lists only openid, email, profile and offline_access; reach is set by the user's role on Drivetrain's Roles page, down to rows. Each token covers one user and one tenant. offline_access allows refresh tokens, whose lifetime Drivetrain doesn't document, so a session can outlast the 24-hour access token. | Drivetrain's requirements list says the person connecting must be a Drivetrain admin, while its ChatGPT steps let members connect as ordinary Drivetrain users. ChatGPT: an admin, who must also be a user on the Drivetrain tenant, creates and publishes the app. Claude: an Owner adds the connector; Drivetrain doesn't say whether non-admins can then use it. |
| Drivetrain MCP server, API key (Claude Desktop via mcp-remote) | A Drivetrain admin copies the key from Settings › API Key, so it may reach the whole account. Drivetrain doesn't document what the key grants, or its expiry or rotation. Its MCP FAQ says there is no static API key, which conflicts with this route; ask Drivetrain. | Drivetrain admin, per Drivetrain. Key from Settings › API Key. |
How do I connect it?
Add https://{tenant}.drivetrain.ai/drive/api/v1/public/atlas/sse to your AI client as a remote server and sign in with Drivetrain; Claude Desktop users can use an API key instead.
- Drivetrain isn't in the Claude directory. An Owner opens Admin settings › Connectors (or Customize › Connectors) › Add custom connector.
- Paste the tenant endpoint as the Remote MCP server URL, leave the OAuth fields empty and select Add.
- Select Connect and sign in with Drivetrain. In the connector's tool settings, set create_board, update_board, add_dataset_transformations and update_dataset_transformations to need approval, or turn them off for read-only use.
Per Drivetrain, this route needs a paid Claude Team or Enterprise plan and Owner rights.
What should a finance team watch for?
Writes are draft-only, so the main risks are what users can read, data that stays in chat history, and the API key route.
- Returned data stays in the user's chat history after their role is reduced, and is visible to anyone the chat is shared with, per Drivetrain.
- Review roles first: sensitive rows such as salaries reach the agent exactly as the user sees them.
- The API key sits in plain text in claude_desktop_config.json, which may be synced. Keep it out of shared folders and version control.
- The API-key route skips per-user sign-in. The key comes from a Drivetrain admin and may reach the whole account; use OAuth where your Claude plan allows, and ask Drivetrain how to rotate or revoke the key.
- Transformation drafts can generate SQL from plain English, per Drivetrain. Check each step before you save the draft.
- Metric names and dataset text are free text. An injected instruction could push the agent to draft a misleading board or transformation, so review every draft.
- Drivetrain logs each tool call, with logs available from Support, and says revoking access is immediate. Revoke in Drivetrain to cut a user off; refresh tokens can outlast the 24-hour access token.
- Drivetrain doesn't document row, rate or query limits for MCP calls.
- Per Drivetrain, clients such as Claude Desktop compute derived figures themselves, for example Revenue per FTE from Revenue and FTE. Check those against Drivetrain's reports.
- Each client tenant has its own endpoint, so a firm needs one connection per client, and Drivetrain lists admin rights as a setup requirement.
- Drivetrain says Anthropic does not train on MCP data. Check Drivetrain's and your AI provider's terms.
How could finance teams and accounting firms use it?
Read tools suit variance questions and reporting; the write tools suit drafting boards and data-prep steps that a person saves.
Inside a company
- Ask for budget versus actual by department and get the figures behind the answer.
- Compare Base Case and Optimistic Scenarios for a metric before a planning meeting.
- Draft a monthly business review board, then edit and save it in Drivetrain.
At an accounting firm
- Summarise a fractional CFO client's actuals against budget before the monthly call.
- Read a client's model structure to explain how a key figure is built.
- Pull a client's Lists or dataset columns to check a mapping before month-end review.
Keep a person approving anything that changes a Drivetrain board or dataset, and treat AI-computed figures as drafts until they tie to Drivetrain's reports.
Want Drivetrain connected to Claude for your team?
We configure scoped connections and a review step before anything changes your books. Fixed fee.
Book a 30-minute setup callChange log
| Date | Change | Source |
|---|---|---|
| 2026-10-09 | MCP overview and tool reference updated: 22 tools (18 read, 4 write), write tools save drafts for review | integrations.drivetrain.ai/extending-drivetrain/using-the-dr |
| 2026-08-31 | Connect guide updated: tenant endpoint, ChatGPT, Claude OAuth and Claude Desktop API-key steps | integrations.drivetrain.ai/extending-drivetrain/using-the-dr |
| 2026 | MCP server product page live, per the dataset (page not reachable on 21 Oct 2026) | www.drivetrain.ai/mcp-server |
FAQ
Is this official?
Yes. Drivetrain hosts the server and documents it in its integrations guide. It isn't in the Claude directory.
Is it free?
Drivetrain publishes no separate MCP price. Per Drivetrain, the Claude OAuth route needs Team or Enterprise; the API-key route works on free Claude.
Does it change data in Drivetrain?
Only through four tools that draft boards and dataset transformations, which a person must save, per Drivetrain.
Is my data used to train models?
Drivetrain says Anthropic processes MCP data under its API policy and doesn't train on it. Its statement doesn't cover ChatGPT; check your AI provider's terms.
Can my accountant use it on my behalf?
Drivetrain doesn't document adviser access. Each person signs in with their own account, so an accountant needs to be a user on your tenant, and Drivetrain lists admin rights as a requirement for whoever sets the connection up.
Sources
- Drivetrain: Using the Drivetrain MCP Server (overview and FAQ)
- Drivetrain: Available MCP tools
- Drivetrain: Connect Common MCP Clients
- Drivetrain: API key (Settings › API Key)
- Drivetrain OAuth protected-resource metadata (scopes)
- Drivetrain: MCP Server product page (not reachable on 21 Oct 2026)
- Node.js download (linked from Drivetrain's guide)
Verified by theFinanceStack, 21 Oct 2026. Vendor list on the CFO Software Market Map courtesy of Jason Staats; AI connectivity data by theFinanceStack. Not tax, compliance or financial advice.