Skip to main content
The Finta MCP server lets AI tools read and update your company’s books. You sign in with your Finta account, so there is no API key to manage.
Agent signup is available while signed out. Supported non-CLI clients can discover and call sign_up with only an email. Coding CLIs use REST signup instead. Existing connected tools still require OAuth.

Tools

  • Onboarding: Answer the company questionnaire, submit it for a decision, connect approved banks, and follow the first import with get_onboarding and update_onboarding.
  • Sheets: Create documents from templates or custom tables with live actuals, forecast formulas and charts. Find documents, read selected values and formulas, search authored cells, edit ranges, tabs and charts, and delete a whole document with the Sheets tools.
  • Reports: Get the income statement by category or the by department, plus the balance sheet and cash flow statement. Get a category’s total, rank its merchants, or break it down by month. Get key metrics month by month, such as cash, runway, revenue growth, gross margin and burn, with list_metrics and get_metric_values; see Key metrics. Generate PDF, CSV or XLSX files with the report export tools.
  • Transactions: Search and get transactions, edit one, or bulk edit transactions. Upload transactions from a bank or card export. Match the two sides of a transfer between your accounts, or unmatch one. Split a transaction between categories, or change or remove a split. Find transactions by tag, or assign or remove a tag on one transaction or many.
  • Invoices: Match one bank deposit to one invoice or a group of 2–100 open invoices, or unmatch a payment. Change the revenue category of an invoice’s line items, open or paid. Invoices are in early access.
  • Tags: List, create, rename, and delete tags. See Tags.
  • Journal entries: Search the debits and credits behind your reports.
  • Accounts and cards: List bank and card accounts and individual corporate cards. Create a manual account to upload to.
  • Reconciliation: See which months or statement periods of each bank and card account still need reconciling, and download each period’s statement PDF. Reconcile one against its statement balance, optionally attaching the statement PDF, or undo the latest one. See Reconciling accounts.
  • Rules: List, create, update, and delete automation rules. Find the rules that use a category, department, tag, account or card by its ID. Rules come back with the names of the category, department and tag they set, and can fill missing transaction tags.
  • Company: Get company details, and list categories, departments, merchants, and integrations. Add a standard category template. Connect or reconnect an integration, and tell Finta once you’ve invited it in the provider’s own app. Departments are in early access.
  • Skills: Get step-by-step help with setup tasks, such as adding Finta to Mercury.
  • Inbox: See what Finta needs from your company, the same items as your Finta inbox. Items close on their own once the work is done.
  • Team: See who is on your team and their access. Invite someone, or change, resend, or revoke a pending invitation.
  • Taxes: Search income and Delaware franchise filings, R&D credits, and vendor 1099 records, or retrieve a known ID with the same tool. Read progress and available documents with the tax tools, or upload an income tax return or Delaware report filed outside Finta.
  • Support: Send Finta support a question, feedback, or bug report.
The existing get_income_statement tool accepts group_by: "category" (the default) or group_by: "department", alongside start_date and end_date. Department reports require the departments feature and use typed expense rows; other sections keep category arrays. The top-level uncategorized_expenses section is available only with category grouping. The tool returns the API response unchanged. See income statement grouping for hierarchy, totals, freshness, and access rules. For a company without departments, list_departments and any tool call that names a department, including one that clears a department with null, return a tool error with the API’s feature_disabled or plan_upgrade_required. For plan_upgrade_required, the tool error includes the company’s Plans link. Retry without the department, and stop offering departments for that company. These tools aren’t marked as early access in the tool list. See Departments are early access. Before choosing a category, use list_categories and read its description: this is the same guidance Finta’s AI categorizer uses, read live and empty when unset. Every category also includes accounting_treatment, which describes automatic bookkeeping that cannot be turned off. Today Equipment is depreciated straight-line over 5 years, with the purchase recorded as an asset and expensed monthly as depreciation; all other categories return null. See Choosing a category. When the user wants a category that doesn’t exist, call list_available_categories. It lists the standard categories the company can still add, each with every category it adds, what each one is for, any accounting_treatment (left out when there is none), and an access value: add, request_required (only Finta adds it; request it with submit_feedback, explaining the intended use) or plan_upgrade_required (the company’s plan doesn’t include it). ChatGPT and Codex show a different value for that plan case, so treat any other value as “the plan doesn’t include it”. Read the descriptions and make sure an option actually fits what the user wants to track before suggesting it, mention any accounting_treatment, then confirm with the user. It needs Accounting setup: View. See Adding a category. To add an option with access: add, call create_category with its key as template. The tool returns created and categories with usable IDs. created is true exactly when the API returns 201: a category was added, or an existing category with the same name that wasn’t yet linked to the option was repaired by linking it. For Equipment, that repair turns on automatic 5-year depreciation. created: false means all of the option’s categories already existed and nothing changed; retries are safe and don’t duplicate categories. It requires Accounting setup: Manage and a connection with write access. A read-only MCP connection returns insufficient_scope. Agents can’t invent categories or choose a parent. For any category that isn’t in list_available_categories, explain that it can’t be added rather than substituting another option. Use bulk_update_transactions with exactly one selector: filters or transaction_ids. Filter jobs have no row cap and transaction_ids takes up to 50,000 IDs. Large jobs take longer: if the tool returns queued or processing, call bulk_update_transactions again with only job_id, backing off between calls until success or failed. Inspect errors even on success; if errors.has_more is true, pass job_id and starting_after: errors.next_cursor to read the next page. A transaction that stops matching the filters or gains a transfer during the job is skipped with transaction_conflict. See Choosing transactions and Reading job errors. Use list_tags to find a tag and reuse it where possible before calling create_tag. At the company’s tag limit, create_tag returns plan_upgrade_required: reuse an existing tag or tell the user which plan they need, and don’t delete tags to make room unless they ask. Listing requires Accounting setup: View; create_tag, update_tag and delete_tag require Accounting setup: Manage. Use update_transaction with tag_id to assign a tag, or null to remove it, one transaction per call. This requires Transactions: Manage. Every call also requires API access: Manage, and writes need a connection with write access. To tag or untag many transactions, call bulk_update_transactions with updates: {"tag_id": "tag_..."}, or {"tag_id": null} to remove. tag_id must be the only update, so a tag change and a category change are two calls. The job doesn’t create tags. A job of up to 5,000 transactions can show the new tag before it saves; see Tag previews. A tag a person set after the job selected the transaction is kept and that transaction is skipped. Removing a tag a transaction doesn’t have counts zero in updated_count. If your client rejects tag_id in bulk_update_transactions, refresh its tool list once. See Tagging in bulk. Before calling delete_tag, explain that the tag comes off every transaction carrying it, rules that only set a tag are deleted, and mixed rules lose their tag action. Other actions and transaction categories are preserved. Confirm these effects with the user before deleting, then refresh list_rules. See Tags. Use create_rule with actions.tag_id for future tagging. Find the ID with list_tags, check list_rules for an existing definition (filter it with tag_id), and confirm the conditions, actions and whether to apply to past transactions. The tool requires an explicit boolean apply_to_existing_transactions; the API defaults it to true. Creation requires Accounting setup: Manage, plus Transactions: Manage for historical application. Tag actions fill only missing tags and never replace one, whatever the transaction’s category, including one another rule set. Historical tagging includes matching past transactions regardless of category or approval, excluding deleted transactions and split lines. Other actions keep their existing eligibility. Use update_rule to change future matching: actions.tag_id sets the tag action, null removes it, and omission preserves it. At least one action must remain. Updates require Accounting setup: Manage and never change past transaction tags. See Automation rules. Use update_transaction with memo to save what the user or a receipt says a transaction was for, so the next person or agent doesn’t have to work it out again. Don’t restate the merchant, amount, date or category. A new memo replaces the old one, so read the current memo with get_transaction first and send the full text; null clears it. get_transaction and search_transactions return memo, which is Finta’s own note and never the bank’s memo text. A memo can be set on approved and reconciled transactions, one transaction per call; bulk updates and rules don’t set it. See Memos. Use submit_feedback to send Finta support a question, feedback, or bug report, when the user asks or on the agent’s own when a Finta tool fails (reported_by: "agent"). It works on read-only connections, and support replies to the user by email. Each call opens a new ticket, so a retry after a timeout may send it twice. Put the problem itself in message. Two optional fields add the context support needs to reproduce it without writing back first: goal (up to 300 characters) is the outcome the user wanted, such as “Split a payroll charge across two departments”, not the call that failed; steps_tried (up to 1,000 characters) lists the Finta tools called, in order, each as <tool>: <what it returned>, with the error code and message for any that failed. Leave out anything that wasn’t a Finta tool call, and never include passwords, API keys, or full account numbers. See Submit feedback. Use list_inbox to see what Finta needs from the company. When an item’s instructions name a skill, offer it, and load it with get_skill only if the user agrees. Otherwise share the item’s url: the user does it in the Finta web app. Items close on their own once the work is done; there is no tool to complete one. Use grant_integration_access once the user confirms they’ve invited Finta in the provider’s own app: Mercury, Ramp, Carta, Pulley, Gusto, or Rippling. Pass the integration’s key from list_integrations, and call it only for a grant_access connection method whose connect_via is provider_app. For Carta and Pulley, also pass url, the company’s page URL at the provider, which you get from the user. Each of these integrations except Pulley has a skill, such as grant-access-carta, that ends with this tool. It returns the integration: Mercury and Ramp read granted and their inbox item closes, Carta and Pulley read connecting until Finta has checked the invite, and Gusto and Rippling read connected. It is marked destructive, so your client asks before running it, and it needs Integrations: Manage and a connection with write access. It replaces complete_inbox_item, which was removed; clients that cached the old tool list need to refresh it. See The inbox and Granting provider access. Use match_invoice_payment with transaction_id and either one invoice_id or an invoice_ids array of 2–100 distinct invoice IDs. The invoice totals must equal the unsplit bank deposit in the same currency. Find candidates with list_invoices using customer and status: "open", then check the sum. Confirm before combining different customers or replacing a categorized deposit’s category. Repeating the exact same set is safe; the group result contains every paid Invoice and its paid_with_invoice_ids peers. To unmatch, pass one invoice_id, its transaction_id, and unmatch: true; do not pass invoice_ids. Explain that this reopens every invoice on the payment, even though the result contains only the requested invoice. Matching and unmatching require Invoices: Manage, Transactions: Manage, and a connection with write access. See Matching one deposit to several invoices. Use update_invoice to change the revenue category or the spread of an invoice’s line items, open or paid, without unmatching it. Read the invoice with get_invoice first, then pass its id and line_items with one entry per line item in the invoice’s order: {"category_id": "cat_..."} to move a line, {"spread_months": 12} to spread it over 12 months (0 recognizes it on the invoice date), both in one entry to do both, or {} to keep it. Use a revenue category from list_categories. Refund lines always stay in Refunds and can’t be spread. It never changes dates or amounts. A paid invoice stays matched and its deposit is rebooked to follow it, so confirm with the user first. It requires Invoices: Manage, plus Transactions: Manage for a paid invoice, and a connection with write access. See Changing line item categories and spreads. Use connect_integration to connect Chase, SVB, Arc, Citi, Bank of America, or American Express through Plaid Hosted Link. Share the returned 30-minute link only with the user: anyone holding it can connect a bank to the company without signing in to Finta. The user signs in to their bank on Plaid’s page; do not automate that page. When they return, call get_integration_connection to check the result. The return page appears on exit too and does not confirm success. If the request is still pending or authorizing, check again later; temporary Plaid errors retry automatically. A late successful session can complete an expired request, so check list_integrations before offering a replacement. Reconnecting an existing Plaid bank remains in Settings > Integrations in the Finta web app. During onboarding, follow the web get-started guidance instead. See Connecting a Plaid bank. Use search_reconciliations with reconciled: false to find periods that still need reconciling, or pass only id to read one period. Before calling reconcile_account, confirm with the user. When a period’s statement_balance_cents is null, ask the user for the ending balance on their statement; no PDF is needed. To keep the statement PDF with the period, upload it with create_file_upload and purpose: "bank_statement" (at most 10 MB), PUT it, and pass the file_id to reconcile_account as statement_file_id. The PDF attaches only if the period reconciles, and replaces any PDF the period had. Each period from search_reconciliations has a statement_file with a download link that works for five minutes, or null. A credit card’s balance owed is positive. If the balances differ, find the cause with search_transactions for that account and period, and leave out a duplicate or an item that clears next period with excluded_transaction_ids rather than deleting transactions. Reconciling locks the period and every earlier open period on the account. unreconcile_account undoes one period per call, newest first, after asking. Reading requires Reconciliation: View; reconciling, undoing, and uploading a bank statement require Reconciliation: Manage and a connection with write access. See Reconciling accounts. Your assistant can only do what your role in Finta allows.

Transaction counts and totals

For “how many / how much” questions, use search_transactions with include_totals: true and limit: 1 instead of paging and adding rows yourself. For all uncategorized transactions, pass:
The response includes total_count and signed total_amount_cents over the whole filtered set, independently of the page size or cursor. Money in is positive, money out negative, and zero matches return zero for both fields. These are full transaction amounts on transaction dates, not accounting-date report totals. See Transaction counts and totals. To total only Uncategorized Expenses or Uncategorized Income, use that category_id from a transaction instead of categorized: false, which selects both. get_category_total still rejects Uncategorized report categories with resource_missing; use search_transactions for these totals. Do not combine include_totals: true with updated_after (400 invalid_request). If the aggregate exceeds its five-second query budget, search_transactions returns totals_timeout as an isError tool result, with no partial result. Narrow filters such as start_date and end_date before retrying. It never substitutes a zero total for an error.

Onboard a company

Call get_onboarding when setting up Finta or after another tool returns onboarding_incomplete. Do not poll it or call it once setup is complete. Re-check only after the user reports an action, asks about status or comes back later. It returns status, saved answers, remaining questions, the eligibility decision, selected integrations, sync progress, and next_step. Do not read it again immediately after update_onboarding, which returns the full object. Tool descriptions stay brief; next_step.message carries the workflow and names the tools to use.
  • Save answers with update_onboarding as the user gives them. Saves are drafts and can be changed until Finta decides, so there is no need to confirm before saving. Banks and cards return options: [], so gather those from context or plain text and map to the accepted input values.
  • Submitting is final. Before sending submit: true, show the user their answers and submit once they confirm.
  • update_onboarding accepts null for every answer field, including enum answers, so clients can replay get_onboarding answers or clear optional answers, for example {"answers":{"incorporation_year":null}} while that follow-up is editable. Omitted answers keep their saved values. Clearing required answers is rejected, and answer locks still apply. The answers object, array items, submit, and notify_when_synced remain non-nullable.
  • Bank/card input enums include canonical provider slugs such as boa and american_express, so schema-enforcing clients can replay answers after recognized “other” names are folded. New answers may use those slugs directly. Saved no_bank_yet and no_credit_cards_yet values are allowed only as unchanged echoes; new selections are rejected. See the accepted values.
  • Answers saved on the web can be replayed unchanged, including saved no-bank/no-card choices. An unchanged echo preserves web ranges, legacy states and additional industries; real changes still obey validation and answer locks.
  • Only connect banks after approval, when the returned step is connect_banks. Follow its message: use connect_integration for new OAuth and Plaid connections, including Chase and SVB. The user approves on the returned provider or Plaid Hosted Link page; then check get_integration_connection. The last selected bank completing either flow starts the import automatically. For a Plaid reconnect, follow the web get-started page instructions in next_step.message, then call update_onboarding with {"answers":{}} to continue setup. This applies before and during the first import: Hosted Link cannot reconnect an existing bank and returns 409 integration_already_connected. OAuth reconnects still use connect_integration.
  • At start_import, every selected bank is connected but importing has not started. Call update_onboarding with {"answers":{}}, then follow the returned step. Reading onboarding alone does not start an import.
  • At wait_for_sync, explain that Finta is importing transactions and setting up the books, and that it can take a while; never call this “sync” or “syncing” in the conversation. Offer the ready email once if not already discussed; set notify_when_synced to the user’s choice. Honor a volunteered preference at any time.
  • If you created the user’s Finta account, tell them once at wait_for_sync to open the environment’s login URL in next_step.message, choose Forgot password, and set a password for later browser access.
  • Do not wait or poll during the import; the ready email lets the user know when the books are ready. Call get_onboarding again only if they ask or come back. sync.check_again_in_seconds is the soonest a later check is useful, not a polling interval. API-started onboarding finishes automatically.
  • At finish_in_web, the web-started import has finished. Ask the user to switch to the company named in next_step.message in Finta first, then open next_step.url, choose See dashboard, and call get_onboarding again. Status stays syncing until that click; the browser may have another company selected.
  • For a failed import, follow next_step.message: reconnect only when reauthentication is required. Use contact_support and the returned email when the connection is still active; a new connection link cannot retry that import.
Both tools need Company: Manage, Integrations: Manage, and API access: Manage. Updates also need a connection with write access. They return the same onboarding object as the API; update errors include the API’s error body, so missing is available after answers_incomplete. The connection stays connected and all tools remain listed while onboarding is incomplete. Outside the onboarding exceptions, calls return onboarding_incomplete as a tool error. Follow onboarding instead of repeatedly retrying another tool. A call-required or not-eligible company is held by the same gate.

Examples

  • “Change our October growth assumption to 8%, keep actuals live, and update the forecast chart.”
  • “Find our cash forecast and show the hiring assumptions alongside their formulas.”
  • “Show our GAAP P&L grouped by department for last quarter.”
  • “Export last quarter’s income statement as a PDF.”
  • “What were our top 10 vendors by spend this year?”
  • “Show our AWS spend by month since January.”
  • “Connect our Chase bank account to Finta.”
  • “Upload this Chase statement to Finta.”
  • “Match the $25,000 transfer from checking to savings.”
  • “Split this $1,000 Amazon order: $600 to Office Supplies and $400 to Software.”
  • “Recategorize every Gusto transaction as Payroll.”
  • “Create a rule that categorizes Cursor charges as Software.”
  • “Create a rule that tags future Vercel charges as Infrastructure.”
  • “Add an R&D category.”
  • “Find transactions tagged Infrastructure.”
  • “Tag this transaction as Infrastructure, reusing the existing tag.”
  • “Tag every Vercel charge this year as Infrastructure.”
  • “What does Finta need from me?”
  • “Have we filed our 2025 taxes?”
  • “What is our 2025 R&D credit, and does payroll submission need action?”
  • “Which 2025 1099 records still need a W-9?”
  • “Invite sam@example.com to Finta as an admin.”

Connect

Every client signs in the same way: log in to Finta, choose a company, and click Allow access.

Signup through an agent

sign_up accepts only the user’s confirmed email and derives app branding from User-Agent. Follow next_step.message: the person clicks Activate in Finta’s one-hour email link in their own browser, then returns to the app to connect and Allow. Never fetch the verification link. New, pending and existing accounts get the same success instructions; existing accounts get no email and can simply connect and sign in. If no account exists but a team invitation is pending, Finta resends the invitation instead of creating an owner account, using the same signup allowance and success response. Follow it to join the invited company. Signup is hidden and refused after connecting. Coding CLIs with a claude-code or codex User-Agent prefix receive 401 plus WWW-Authenticate and use REST signup instead. If connecting requires a restart, explain how to return to this chat and call no Finta tool until the connection has loaded. Once connected, follow the response’s onboarding guidance and next_step.message. For other signed-out clients, protected calls receive HTTP 401 with WWW-Authenticate. openai-mcp instead gets HTTP 200 with a JSON-RPC tool error, isError: true, “Sign in to Finta to continue.” and _meta["mcp/www_authenticate"] containing the Bearer resource-metadata challenge. Invalid, expired or wrong-resource bearer tokens and API keys are still refused. Manual claude.ai and ChatGPT signup/challenge verification is outstanding. Anonymous discovery/signup uses a separate 6,000/minute per-IP request bucket, pending confirmation of the shared-provider request budget. Signup shares email allowances with web and REST signup. For signup verification emails only, matching Anthropic or OpenAI connector IPs use one shared bucket per provider: 100 attempts per 10 minutes and 1,000 per 24 hours across all matching IPs and recipients. Recipient limits are unchanged, and other auth email purposes keep the ordinary per-IP limits even on provider IPs. The published ranges are pinned in code; User-Agent does not select the higher allowance. Signup consumes no monthly MCP allowance. Signup errors are tool results with isError: true, the REST code in text and structured content, and retry seconds for throttling.

Claude

  1. Open Customize > Connectors in the Claude desktop app or on claude.ai.
  2. Click + Add, then Add custom connector.
  3. Enter Finta as the name and https://mcp.finta.com/mcp as the URL.
  4. Click Continue, keep the default sign-in settings, then click Add and sign in.
On a Team or Enterprise plan, an Owner first adds Finta under Organization settings > Connectors. Members then click Connect on Finta in Customize > Connectors.

ChatGPT

  1. Go to ChatGPT Plugins.
  2. Click Add, then Create custom MCP server.
  3. Enter Finta as the name, Startup taxes, accounting, and finances as the description, and https://mcp.finta.com/mcp as the connection. Choose OAuth for authentication.
  4. Click Create as a plugin and sign in.

Claude Code

Then run /mcp, select finta, and choose Authenticate.

Codex

Requires Codex 0.147.0 or newer.
Codex opens Finta in your browser to sign in. To sign in again later, run codex mcp login finta.

Other clients

Add https://mcp.finta.com/mcp as a remote MCP server over Streamable HTTP. Finta supports OAuth with Client ID Metadata Documents and Dynamic Client Registration, so there is nothing to register first. Request the finta:read and finta:write scopes; without finta:write, tools that change data are unavailable. In Finta, Settings > MCP also has a prompt you can paste into Claude Code, Codex, or Cursor to set up the connection for you.

Limits

After onboarding, tool calls count toward a monthly allowance shared by your company: 10 on Formation, 300 on Startup, and 30,000 on Growth. Finta’s internally enabled Enterprise entitlement bypasses commercial monthly API/MCP quotas; calls are still counted and permissions and abuse limits still apply. It is separate from the API’s monthly limit and resets on the 1st of each month. Before the company is onboarded, no tool calls count toward that allowance or fail for reaching it. get_onboarding, update_onboarding and submit_feedback are always free, including after setup or at the cap. Calls are still logged with counted: false; permissions, billing and security checks still apply.

Billing invoice refusals

An overdue unpaid billing invoice explicitly linked by Finta to an enabled Enterprise product pauses every tool call, independently of the company’s subscription or plan. The invoice payment rules use Pacific due dates with no payment-processing grace period. Pay overdue linked invoices at app.finta.com or contact support@finta.com; reconnecting or repeatedly calling tools does not resolve this. The connection stays connected: initialize and tools/list still work. All tool calls are refused before subscription and monthly allowance checks, including writes, usage tools and tools that otherwise make no API request. A refusal is an HTTP 200 JSON-RPC tool result with isError: true and exactly one text content item containing the API’s human-readable message. There is no error-code field or structuredContent; do not expect the API error envelope or parse the message as a stable contract. The overdue refusal is logged as 403 and is not counted against the monthly allowance. If payment status cannot be verified after the cutoff, the same text-only error result is logged as 503 and remains uncounted. The HTTP response still uses status 200 but includes Retry-After: 60. Wait that interval, then use bounded exponential backoff if verification remains unavailable. A Stripe outage before the cutoff does not block access for this reason. See invoice_status_unavailable.