What is Sure?
Sure is a self-hosted personal finance app: accounts, transactions, budgets, net worth, investments, loans and bills, in a Rails web app with a macOS desktop shell and a Flutter mobile companion. It is a community fork of Maybe Finance, whose team spent close to $1 million building the app before archiving it in July 2025; Sure picked up the code that month under the same AGPL-3.0 licence and has shipped a release every month since.
The AI sits in two separate pipelines. A chat assistant answers questions about your accounts and transactions by calling Sure's own functions, and background jobs categorise transactions and detect merchants. Both use any OpenAI-compatible endpoint (OpenAI's gpt-4.1 by default), Anthropic natively, Gemini, or a local model through Ollama or LM Studio. A Model Context Protocol endpoint lets Claude, GPT agents or your own agent read and change Sure data, and an mcp_only mode removes the in-app chat for people who bring their own agent.
Bank connections come through Plaid, TrueLayer, Enable Banking and regional providers such as Monobank and FIO, plus crypto wallets and CSV or PDF imports. The project's site at sure.am runs a hosted instance with sample data; its pricing is not published.
Is Sure worth it?
Sure's best argument is that it survived. Maybe Finance had about 54,000 GitHub stars when it was archived; the fork now has about 14,400 and has kept a monthly release cadence for more than a year, most recently 0.7.6 on 31 October 2026 and a hotfix on 12 November. That is unusual for a rescued codebase, and it matters for software that holds your financial records.
The AI design is more careful than most. The chat assistant and the categorisation jobs are configured separately, so you can point categorisation at a small local model and chat at a hosted one, or turn the in-app assistant off and connect an external agent over MCP. The docs are frank about the hardware: 8 GB of VRAM runs a 7B model for basic chat, 16 GB or more is recommended, and CPU-only inference is 10 to 100 times slower. By default one worker thread handles all AI work, so a long PDF import waits behind chat, and categorisation sends batches of 25 transactions per model call. Version 0.7.6 added a Bayes classifier that categorises without spending tokens at all.
The security model needs attention. The MCP endpoint grants a read_write scope over OAuth 2.0, or a static bearer token tied to one user's email, and the docs warn that its tools can import statements and change transactions, budgets and categories. Encryption keys for stored provider credentials are derived from SECRET_KEY_BASE unless set explicitly; the Compose file now requires your own value after earlier example files shipped a default, and a key-rotation guide covers anyone affected. Version 0.7.6 also turned on anonymous PostHog product monitoring, which self-hosters switch off with POSTHOG_FEEDBACK_ENABLED=false.
The rough edges are real. 916 issues and pull requests are open, Plaid's production access is unavailable to European users and needs the instance on a public domain with SSL, the mobile app covers sign-in, balances and read-only budgets, and the Qdrant document-search backend is scaffolded but cannot store documents. A Rails app with Postgres, Redis and a Sidekiq worker is also more to keep running than a single binary, which is the price of the assistant and agent endpoint.
Where it earns its place
A privacy-minded household running a home server with a 16 GB GPU
Ask questions about two years of spending without sending transactions to a cloud model
Run Sure with Docker Compose, point the LLM provider at Ollama with a 13B or 14B model as the docs recommend, and set categorisation to the token-free Bayes classifier. The only cost is the hardware; CPU-only inference is 10 to 100 times slower.
A freelance developer who already uses Claude Desktop daily
Let an agent answer "what did I spend on software this quarter" and tidy categories
Set Sure's assistant to mcp_only and connect Claude through the OAuth MCP endpoint, which grants read_write access. Model costs fall on your Claude plan; because the tools can change transactions and budgets, review what the agent edits.
A small community group in the US that hosts shared services on Render
Give a few members a hosted finance app with bank sync and no AI bills
Deploy the Sure - No AI Render blueprint, which provisions the web app, Sidekiq worker, Postgres and key-value store with AI tokens blank, then add Plaid keys on a domain with SSL. Plaid access can take more than 24 hours to approve.
Before you commit
- Setup
- Docker Compose following docs/hosting/docker.md: generate a SECRET_KEY_BASE (there is no built-in default), then open port 3000. One-click options exist for Render, PikaPods and Hostim. AI needs an OPENAI_ACCESS_TOKEN, ANTHROPIC_API_KEY or local endpoint, plus consent in the UI.
- Integrations
- Plaid, TrueLayer, Monobank, FIO, SnapTrade, Enable Banking, Kraken and on-chain wallets; OpenAI-compatible APIs, Anthropic, Gemini, OpenRouter and Ollama; MCP for external agents; an OpenAPI-described HTTP API; optional Langfuse tracing and pgvector or OpenAI for document search.
- Data and privacy
- Self-hosted in your own Postgres. Provider credentials are encrypted with keys derived from SECRET_KEY_BASE. Prompts go to whichever model provider you configure, or nowhere with a local model. Anonymous PostHog product monitoring is on by default; POSTHOG_FEEDBACK_ENABLED=false turns it off. AGPL-3.0 requires sharing changes if you offer a modified version as a service.
- Cost at scale
- Software is free. A hosted model bills per token at the provider's rate; the docs track prices in config/llm_pricing.yml and show usage in the app. Local inference needs 8 GB of VRAM at minimum and 16 GB recommended.
- Limits
- One AI worker thread by default, so background jobs queue behind chat; Plaid unavailable in Europe; mobile app limited to balances and read-only budgets; Qdrant backend not functional; 916 open issues and pull requests.
What’s good
- Monthly releases for more than a year since the fork
- Local models via Ollama for chat and categorisation
- MCP endpoint with OAuth and an mcp_only mode
- Token-free Bayes categoriser as of 0.7.6
What’s not
- MCP tools can write to your financial data
- Product monitoring on by default since 0.7.6
- Plaid unavailable to European users
- 916 open issues and pull requests; thin mobile app
Best for
Self-hosters who want a full personal finance app with an AI assistant they can run on a local model or reach from their own agent over MCP.
Sure: common questions
- Is Sure worth it?
- Sure is the self-hosted finance app that takes AI seriously without making you send your bank history to OpenAI, provided you are willing to run a Rails stack and read the docs. We rate it 3.9 out of 5.
- How much does Sure cost?
- Free and open source under AGPL-3.0 for self-hosting. A hosted instance runs at app.sure.am; no price is published. AI features bill at your model provider's rates unless you run a local model.
- What are the downsides of Sure?
- MCP tools can write to your financial data; product monitoring on by default since 0.7.6; plaid unavailable to European users; 916 open issues and pull requests; thin mobile app.
- Who is Sure best for?
- Self-hosters who want a full personal finance app with an AI assistant they can run on a local model or reach from their own agent over MCP.
- How much does Sure cost for a team or in production?
- Software is free. A hosted model bills per token at the provider's rate; the docs track prices in config/llm_pricing.yml and show usage in the app. Local inference needs 8 GB of VRAM at minimum and 16 GB recommended.
- Is Sure safe for company data?
- Self-hosted in your own Postgres. Provider credentials are encrypted with keys derived from SECRET_KEY_BASE. Prompts go to whichever model provider you configure, or nowhere with a local model. Anonymous PostHog product monitoring is on by default; POSTHOG_FEEDBACK_ENABLED=false turns it off. AGPL-3.0 requires sharing changes if you offer a modified version as a service.
Answers checked Nov 13, 2026.