Your PC does the work. You get the proof.
Sonjari is a free local AI agent that works right on your own PC, on a model running on your GPU. It researches the web in your real Chrome, builds Excel and Word documents, and reads and fixes code, and it shows you what it actually read. It works on its own, takes on tasks handed off by Claude Code, Codex or Cursor, and can take work from your phone over Discord. No API keys, no subscription, no usage meter.
- No meter on anything that runs on your PC
- Picks up when Claude Code or Codex hits its limit
- Every run leaves receipts you can check
Windows 10/11 · NVIDIA GPU · Free on the Microsoft Store
교보문고에서 '파이썬' 책 3권을 찾아 제목·판매가·적립 포인트를 엑셀로 정리해 줘 “Find 3 books for ‘Python’ on Kyobo Book Centre and put title, price and reward points in a spreadsheet.” Redrawn from the record of a real run · 28 Sep 2026, 23:51 KST · one RTX 3090 PC · 9 steps, 0 failed · run 026df5bd
One sentence in. A real file out.
With Claude Code: it hands the research to Sonjari
Same task, in the real app
The answer on the left, the file it made, and the requirement check at 3/3. On the right, Sonjari’s computer: the step ruler and the page it read last.
What we promise, and how it's going
A promise without a record is a slogan. Each one below has the number we measured, how many runs, and when. When we break one, it stays on this page.
-
It won't take your mouse.
KeptChrome used to jump in front of your work when a task started. It did in 7 of 8 runs. After the fix, 0 of 8.
n=8 · 2026-09-21 -
It says so when it couldn't do it.
Broken onceHonesty bench: asked across 3 routes with 3 questions, it pretended to have read something 0 times out of 9.
n=9 · 2026-09-3028 Sep 2026: in one run it reported a task as done after copying a recommendation banner, with no warning. We fixed the checker the same night. The rematch that night: 0 false completions in 6 runs.
-
No meter on what runs on your PC.
KeptThe model runs on your GPU. Web search runs in your own Chrome, so there is no search API key and nothing to top up. We don't spend your Claude or ChatGPT subscription behind your back.
How it's built, not a benchmark
It works next to the AI you already pay for
When Claude hits its limit, keep going
Tell Sonjari you're out. It finds where Claude Code or Codex stopped, puts that work on a new branch and carries on locally.
Spotting the stop: caught 52 of 52 stopped sessions, 1 false alarm in 458 ordinary messages. How often the handed-over work gets finished: not measured yet.
Hands for your frontier model
Connect it to Claude Code, Codex, Cursor or the Claude app over MCP. The frontier model plans; Sonjari does the reading and clicking on your PC and sends back the result.
On Korean shopping sites, Claude Code's own cost was lower with Sonjari in 9 of 9 paired runs ($1.71 vs $3.69 alone), and correct rows did not drop. It took about 4× longer. 27 rows per arm, 2026-10-15.
Your files stay on your disk
Excel, Word and HWP files are read and written on your PC. Nothing leaves it unless you send it, and a bug report shows you exactly what it will send first.
The run above wrote its spreadsheet to your Documents folder and re-checked it: 4 rows.
For developers
How to plug it into your own workflow. All on GitHub.
What it can't do yet
- Windows only. Windows 10 or 11.
- It needs an NVIDIA GPU. 6 GB runs a small 4B model that works best next to the AI you already use. 12 GB runs a 9B, and 24 GB runs the 27B we test on. AMD and Intel cards are not measured yet.
- It is slower than the cloud. On the same price research, Claude Code alone took 96 s and with Sonjari 401 s (median). On one coding task Hermes Agent was about 2× faster.
- The English interface is new. It shipped in October 2026. Tell us where the wording is off.
- The numbers are still small. On English sites a 12 GB-class model finished GitHub trending in 7 of 7 runs and arXiv in 4 of 4. One of those runs put a wrong number in the table.
- This is the first public release. Expect rough edges and tell us about them.
Every run we measured, including the ones we lost
Newest first. Each line says what we tried, how many runs, and what happened.
Try it on your PC.
It installs in one click from the Microsoft Store and updates itself. The source is on GitHub under Apache-2.0. We read every report.