Why it mattersA team spread across four coding agents can now back up its conversation history in one place, feed it into an evaluation harness, or hand a specific session to another agent, without the vendor's cloud in the loop.
The developer kruzovic7 published ai-data-extractor, a local tool that walks the disk for chat histories left behind by coding assistants and normalises them into one JSONL format. The repository is MIT-licensed, was created on 11 September 2026, and has picked up 825 stars and 132 forks in five days.
What it reads
The tool auto-detects installed assistants on Windows, macOS and Linux and reads whichever it finds. The README lists Claude Code, Cursor, Windsurf, Aider, Cline and Roo Code, and says "and eight others". It handles the storage each one uses, which differs by vendor: SQLite databases for some, JSONL files for others, and nested JSON trees for the ones that write per-session files. Every conversation comes out as the same JSONL record: user messages, assistant responses, code blocks, diffs, tool calls, timestamps and the model name where the source recorded it.
Why 825 stars in five days
A team that uses two or three of these assistants a day has no way to look at all its recorded work in one place. Each vendor keeps the transcript in its own location and its own schema, some of them in SQLite files a person cannot open with a text editor. If a team wants to back the sessions up, feed them into an evaluation harness, or move a promising conversation from one agent to another, it has to write the code for each vendor separately. This tool does that work once, and the JSONL it produces is the format most evaluation and fine-tuning stacks already expect.
The star count came fast enough that a check is worth showing: the repository was created on 11 September 2026 at 18:12 UTC per the GitHub API, has 825 stars, 132 forks, and is 59 KB in size. That is a real project with real reach for its age, not a promoted empty repository.
What to think about before adopting it
Chat histories may contain more than the model output. Some of these assistants log the working directory the agent was pointed at, the file paths it read, and in a few cases the file contents themselves. A team that feeds this JSONL into a shared evaluation harness or a fine-tuning job is copying its own source code into that job by default. The tool preserves this because the point is completeness, so the pipe from extractor to next step is the place to redact, not the extractor.
The other adoption cost is trust in the schema. This is a one-week-old project with 6 commits from a single author. The current output is fine for a backup or a one-off analysis, and worth reading before wiring it into a production evaluation loop.
Source
kruzovic7/ai-data-extractor on GitHub, MIT license.
Source: kruzovic7 on GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.
