Local AI time tracker: on-device models, no cloud, no account
A local AI time tracker records and matches your work on your own computer. How Planim runs its models on-device, where the file is, and its full network list.
Contents
A local AI time tracker does the recording and the AI matching on your own computer, so window titles, URLs and file names never reach a vendor's server. Planim works this way. It records the day into a SQLite file on your machine, runs two language models on your CPU to match each block of work to a task, and has no account to sign up for.
I make Planim, and this page shows what each part of it does on your computer, down to the folder and the network connections.
What "local" has to cover
A tracker has four parts that can each be local or remote, and a vendor's "local" usually describes only one of them. In Planim all four are local:
| Part | In Planim |
|---|---|
| Capture | On the device |
| Storage | A SQLite file on your disk |
| The AI | Two models on your CPU |
| Account | None |
Other products are local in part. Clockify's Auto tracker keeps its record on your computer for 45 days, and turning it into entries is manual. Memtime and ActivityWatch also keep everything on the machine; Memtime sorts the time by rules you write, and ActivityWatch does not sort it. The two with AI run it remotely: Timely in its cloud, on the activity its Memory app uploads, and Rize on its servers after filtering on the device. Those facts are from each vendor's own pages, checked November 2026.
The on-device AI models
Matching activity to a task means reading text. "grid-template-areas - CSS | MDN" in a browser tab and a task called "Category ribbon on the timeline" share no words, and a person can still tell they belong together. Two small models do that job in Planim.
The first is an embedding model. It turns the words of each block of work and the summary of each open task into vectors and finds the tasks closest in meaning. It returns the ten best candidates above a similarity floor.
The second is a reranker. It reads the block against each of the top five candidates as a pair and reorders them. It is slower and more exact, which is why it only sees five.
Both run on the CPU during analysis, and nothing is sent to an inference API. Their answers are two of ten votes; the other eight come from issue keys, known URLs, file paths, your past verdicts and what else was open at the time. What the AI in an AI time tracker does covers how the votes are merged and how confidence is calibrated.
Learning is local too. Your verdicts go into a journal in the same database, and the channel weights are retrained on your machine from that journal.
Where a local time tracker keeps the data
On macOS: ~/Library/Application Support/dev.auto-time-tracker/. On Windows: %LOCALAPPDATA%\dev.auto-time-tracker\. It is a SQLite database, and any SQLite tool opens it and shows every row.
Credentials for your connected trackers inside the file are encrypted with AES-256-GCM. The key is held in the macOS Keychain or Windows Credential Manager, so a copy of the file on its own contains no working token.
Raw events are pruned after 90 days; 30, 60, 180 and 365 are selectable. Export writes every table as JSONL with CSV views of episodes and drafts. Delete everything wipes the database and the stored credentials.
The complete network list
- Your trackers. The app reads your task lists so it has candidates to match against, and it writes worklogs to Jira, time entries or comments to Asana and comments to Trello. Writing happens only when you press Push worklog, and only for drafts you confirmed.
- One model download. The two models are fetched once from Hugging Face, pinned to a specific version and checked against a hash.
- Sign-in. Connecting a tracker with the sign-in button goes through our relay, so the app does not ship client secrets. The relay holds the session for at most ten minutes and deletes the tokens the moment the app collects them. Pasting an API token skips the relay.
There is no telemetry, crash reporter or analytics. Capture, analysis and review run on the machine, and the three items above are the only reasons the app opens a connection.
Why on-device matters at work
Client contracts. Window titles and file names can count as client material under an NDA. A log that never leaves the laptop gives a security review no activity data to assess.
Works councils and GDPR. In Germany, § 87(1) no. 6 BetrVG gives the works council a vote on technical systems able to monitor performance, and GDPR Article 5(1)(c) limits collection to what is necessary. With Planim the employer receives worklogs and nothing else: issue, date, hours, authored by the employee. This is not legal advice.
Your own preference. The record of a working day includes the bank tab and the message from your doctor. Denylists drop chosen domains and apps at the moment of capture and purge what was stored for them, and private windows in Chromium browsers and Firefox are skipped. More on that in a time tracker without screenshots.
How to verify a "local" claim
- Find the data folder and open the file.
- Run the app with the network off and see what still works.
- Watch its connections with Little Snitch on macOS or Resource Monitor on Windows and compare them with the vendor's list.
- Ask where the model runs, and which text is sent to it.
- Look for a sign-up form. An account usually means a copy of your data on the vendor's servers.
The folder path and the network list above are what to check Planim against.
Planim is in development for macOS 14 and Windows 10/11. Leave an address below to get the local automatic time tracker when it ships. For how the category works in general, start with how an automatic time tracker works.