Image: GitHub
Why it mattersA team that has run several coding assistants side by side can now put every conversation into one file and compare or fine-tune on it, rather than losing the record inside each vendor's own database.
The developer bawadou published ai-data-extractor, a Python tool that reads the local chat history of ten AI coding assistants and writes it out as one shared JSONL file. The repository was created in mid-August and has gained 552 stars and 102 forks.
Which tools it reads
The list covers most of the tools a working developer has probably tried in the last year: Claude Code, Codex CLI, Cursor, Windsurf, Trae, Continue, Gemini CLI, OpenCode, Cline and Roo Code, and Aider. Each of those keeps its transcripts in a different place, in a different shape, on the local disk. The extractor knows where to look on macOS, Linux, and Windows for each one, and reads the files without opening them for writing.
What comes out
Every conversation is written as one line of JSON in a single JSONL file. The line records the messages, the code context the assistant saw, the diffs it proposed, the tool calls it made, timestamps, a session identifier, and the project path where the conversation happened, whenever the source tool records that information at all. Fields that a given tool does not track are simply left off, so the format still works when a source records less.
Local, and read-only by design
The extractor runs on the same machine that has the transcripts, and never opens any of them for writing. The README states plainly that the tool never writes back into a source database. The author also warns that transcripts often contain code, tokens, and other secrets, and asks users to scan the extracted file for those before sharing it, and to check each vendor's terms of service before using an export for training.
A developer who has used more than one coding assistant this year has, without noticing, spread months of their working history across several private databases in several formats. Some of that history is disposable and some of it is the closest thing the team has to a record of how a hard change actually got made. Until now, extracting any of it was a per-tool exercise: read the SQLite schema, guess the JSON structure, hope the timestamps match.
The existing exporters for this are all vendor-specific. ai-data-extractor offers one output format that behaves the same across ten tools, licensed MIT, so a team can do the actual work with it: build an internal search over old sessions, train a private fine-tuned model on real code review turns, or move a project's history from one assistant to the next without losing what the last one saw. The 552 stars in about two weeks are proof that a lot of people had the same problem of scattered, disorganised history and were waiting for a way to solve it once.
Source
Primary source: bawadou/ai-data-extractor 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 using them to release software. Short, and only when there is something worth reading.