Export the code context you actually need
A coding conversation can contain the fix you want to keep and the credential you should never share. AI Data Extraction collects local assistant histories into JSONL, retaining code context and tool output alongside the conversation. That makes it useful for recovering your own work across Cursor, Claude Code, Codex, and other supported tools. It also makes the output more sensitive than a chat transcript copied by hand.
The 0.2 MB checkout we measured is a small starting point for a concrete job: turn several assistants' storage layouts into files your own scripts can inspect. Each conversation occupies a JSON line, with source and message metadata. You can combine exports or select conversations containing diffs. The value is preserving evidence of how work happened, provided you can handle everything else that evidence contains.
What happened when we ran it
Our sandbox installed commit 2b65f74 in 11 seconds, adding 35 packages and occupying 37 MB on disk. The build succeeded in 2 seconds. Pytest then passed all 11 tests in 7 seconds, with 0 failures. We ran that snapshot on October 3, 2026, in an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Python 3.12, and no secrets.
Pip-audit reported 0 known vulnerabilities. The checkout had 17 files and about 4,042 source lines, a tests directory, no Dockerfile, and 0 CI workflow files. Those are encouraging repository checks for a utility you may run against private developer history. They don't establish extraction completeness across every assistant or storage revision. The supplied run contains no accuracy benchmark against a real personal archive.
Local history is the input, and JSONL is the output
The README's basic extraction route needs Python and readable assistant storage, without a hosted model credential. python3 extract_claude_code.py searches Claude directories, while python3 extract_codex.py reads rollout files. Cursor extraction handles SQLite stores rather than assuming every assistant writes plain chat logs. Timestamped files go into extracted_data/. That is an approachable interface for someone comfortable checking paths and reading structured output.
Our 37 MB installed environment should not be confused with the space needed for your archive. Exported code, diffs, and tool results add their own volume. Discovery also depends on where the assistant actually stores history: the troubleshooting guide tells users with custom locations to edit the search paths. On macOS, access permissions may intervene. SQLite locking can require closing the assistant or using a readable copy.
Raw exports can carry secrets and proprietary code
The README explicitly warns that output can include API keys, personal paths, and proprietary code. That's a reason for a company developer to stop before exporting an employer's sessions into a personal training collection. Ownership of the laptop does not settle permission to reuse every file fragment or tool result. The toolkit's stated purpose is extracting your own data, and its disclaimer leaves rights and handling decisions with you.
A green 11-test run does not certify a clean corpus. The optional filter_privacy.py processes nested string values locally and writes separate outputs, with overwrite protection. It needs its own runtime and model assets. The documentation recommends reviewing the result and adding secret scanning. I would keep original exports restricted, inspect the filtered files, and avoid sending either collection to a remote model by default.
Generating skills adds a model service to the workflow
corpus_to_skills.py samples conversations and asks an OpenAI-compatible model to produce reusable Agent Skills. Its default endpoint is localhost, and you choose a model through a command option or environment setting. Output becomes directories containing SKILL.md, with checks on names and response structure. This is a useful way to draft procedures from repeated work, especially when you want to recover decisions instead of retaining every transcript.
The 11-second install measured in our sandbox does not establish that a model service is ready or that a generated skill is safe to install. A remote endpoint changes where sampled history travels. The script treats corpus text as untrusted evidence, but its prompt cannot guarantee that output omits sensitive material. Read each generated procedure, verify its commands, and remove details that belong only to the original project.
October changes make the measured commit a dated baseline
GitHub reported a last push of October 7, 2026, with 1,346 stars and 7 open issues and PRs. The original owner URL redirects to sybil-solutions/ai-data-extraction. Recent merged work consolidated Pi and OMP parsing and added Droid extraction; an open contribution requests Devin Desktop support. These are signs of ongoing adaptation to assistant storage, rather than grounds to assume every newer format already works. The latest-release endpoint returned no release.
The October 3 snapshot had 0 CI workflows, while the current tree includes a test workflow. That distinction matters: our passing checks describe commit 2b65f74, not every subsequent change. The README calls the project production ready, but I would pin the commit and compare a representative session with its export before trusting an archive migration. Missing tool output can matter as much as a missing answer.
A narrower exporter is better for readable transcripts
Claude Conversation Extractor targets readable Markdown exports from Claude Code, which is preferable if human reading is the whole job. Simon Willison's LLM stores prompts and responses from its own CLI in SQLite, making it an alternative for collecting future interactions under a known logging system. Neither choice replaces this toolkit's cross-assistant recovery task. Choose based on where the history comes from and what you'll do with it next.
AI Data Extraction earns a trial because our 35-package setup stayed small and all 11 tests passed. Its useful boundary is equally clear: extracting a corpus and approving that corpus for reuse are separate jobs. Before your first bulk export, pick one conversation with a code edit and a tool result. Check that both survive, then decide who should be allowed to read the resulting file.

