For AI tooling · in development
Your assistant is fluent at arithmetic it gets wrong.
An MCP server exposing the checker behind math.photos as a tool: read math out of an image, verify a chain of working line by line, and return which step broke. So the model can check itself instead of asserting confidently.
No account, no card. One email when there’s something to show you.
The failure this addresses
Language models are good at producing something that looks like a derivation. They are noticeably worse at noticing when their own line 6 doesn't follow from line 5 — and worst of all at admitting it, because the failure is fluent.
Giving the model a tool that independently checks working turns an assertion into something verifiable. The check runs on a different engine with a different job, which is the entire point: a model grading its own homework is not a check.
Tools we'd expose
read_math— image in, structured math out. Handwriting and screenshots included, with the transcription returned separately so the model can see what was actually read.check_working— a problem plus a sequence of steps; returns valid/invalid, the first bad line, and the rule that was broken.solve— worked solution with steps, for when there's nothing to check yet.
Where this would actually get used
- Assistants handling homework help. The screenshot arrives in chat already — we see the queries. Right now the model guesses at the image and guesses at the algebra.
- Agentic workflows touching numbers — anything doing financial modelling, unit conversion or derivation as a step, where a silent arithmetic error propagates.
- Eval harnesses. A step-level checker is a better scoring function than exact-match on a final answer, which rewards a wrong method that lands on the right number.
Open questions
- Hosted or local? Remote server is simplest; some of you will want it in your own environment, and images are the reason.
- Auth and metering. Per-key usage is the obvious model, but free-tier shape matters more for a tool than for a product.
- Which client first? Tell us in the note — it decides what we test against.
Related: if you want this as a plain HTTP endpoint rather than over MCP, that's the API page, and it's a separate list.
Not shipped yet. Interestingly, we're already being cited by assistants without one — Bing's AI reporting shows 859 Copilot citations to our pages in three months, and real users typing things like “can you do the math based on my screenshot to make sure that's true?” into chat. The tool call seems like the obvious next step, so we're checking whether anyone would actually wire it up.