AI insights
An AI time tracker that answers questions about your own week
The useful application of a model here is not tracking, it is reading. The tracking is arithmetic. Turning a thousand rows into a sentence you can act on is the part worth automating.
In short
An AI time tracker applies a language model to time you have already tracked: summarising the day, answering plain-language questions about your hours, and flagging workload patterns. In Stunda the tracking itself is not done by a model. Scoring and lookups run as ordinary code against your own rows, and the model writes the sentence around the result.
- Digest, cached per day
- 1/day
- Workload scoring window
- 28 d
- History before it will read workload
- ~14 d
- MCP access, every plan
- Free
The three AI surfaces, and what each one actually does
These are the features in the product, described as they are built. There is no fourth one coming in a banner.
- A daily digest that reads your day back to you
- Written from the day's own sessions: what took the time, what fragmented, and what is billable and not yet on an invoice. Cached per day, so opening it twice costs nothing.
- Questions in plain language, answered from your own rows
- Ask how much unbilled work is sitting in last week and the question is resolved into a typed intent, then into an access-scoped lookup. The model never writes the query, which is why it cannot reach a row your account cannot.
- Workload signals that refuse to guess
- Pattern scoring runs in ordinary code over your last 28 days; the model writes the sentence around the result. Under about two weeks of history it declines to give a reading rather than inventing one.
The model does not write the query
This is a design decision worth being explicit about, because the alternative is common and the difference is invisible from the outside.
The easy way to build a natural-language interface over a database is to let the model generate the query. It works impressively in a demo and it means the boundary of what the model can read is a prompt instruction, which is not a boundary. Stunda resolves your question into one of a fixed set of typed intents, and the lookup for that intent runs with your account's own access scope applied. A question that would reach data you cannot see does not return the wrong answer, it does not run.
It refuses to give you a reading it cannot support
The workload feature is the most credible thing in this section precisely because of what it will not do. With under about two weeks of history it says it does not have enough to tell you anything, instead of producing a confident sentence about your burnout risk from four days of data.
That is not a limitation waiting to be removed. A tool that always has an answer is a tool whose answers mean nothing, and an assessment of how hard someone is working is the last place to start bluffing.
Your tracked time, inside the AI tools you already use
Stunda hosts a Model Context Protocol endpoint, so Claude, Cursor, VS Code and Codex can read your time directly.
- Hosted, not self-run
- Stunda serves the endpoint at api.stunda.io/mcp/, so there is nothing to install, clone or keep running.
- Version negotiated per request
- It serves the 2026-07-28 protocol revision and every earlier one from the same URL, so clients negotiate per request and you never pin a version.
- 5 read tools, exactly one write
- The reads are annotated read-only so your client can auto-approve them and stop to ask only about the write.
- Runs strictly as you
- Authenticated with a personal access token you create and revoke yourself, stored hashed. Every call carries your own access scope.
Questions
What is an AI time tracker?
A time tracker that applies a language model to hours you have already recorded: summarising your day, answering questions about your own data in plain language, and surfacing patterns across weeks. The measurement itself is arithmetic and does not need a model.
Does the AI decide what my time was spent on?
No. Project assignment comes from evidence the tracker captured and from your own corrections. The model reads the result and describes it.
Can the AI see other people in my workspace?
Only what your account can see. Questions are resolved into typed lookups that run with your access scope applied, rather than the model composing a query, so it is not possible for it to reach a row you could not open yourself.
Can I use my own AI client with Stunda?
Yes. Stunda hosts an MCP endpoint that Claude Code, Claude Desktop, Cursor, Codex and VS Code can connect to. It serves 5 read tools and one write tool, and it works on every plan including Free.
Does using the AI features cost extra?
The digest and the assistant are part of the paid plan. MCP access is on every plan including Free, and it does not consume any AI quota, because the reasoning happens in the client you are already paying for.