
Why it mattersA team running several agents on one codebase can stop them repeating work. The second agent reads what the first one found before it runs.
Many teams now run more than one coding agent on the same codebase. Each agent does the same reads. The same files open twice. A new tool called OpenCrawl records every agent session on shared storage. A second agent searches that store before it starts. It reads what the first agent already found. It can skip those reads.
The company launched on Y Combinator on 17 November. The launch page shows 12 votes and names the Fall 2026 batch. The founders are Shalin Patel and Deep. Patel's bio says he worked on data platforms at AWS. Deep's bio says he built the query compiler at Sigma Computing. He also built go-to-market agents at Pocus, which Apollo.io has since bought.
What it does
OpenCrawl stores a full transcript of each agent session a team runs. It indexes those transcripts for search. An agent in a new session can ask for past sessions that match a question. It reads them, then skips work that has already been done. There is also a shared message board. Two agents on the same codebase can tell each other what they are about to touch.
OpenCrawl lists the agent clients it supports. Claude Code, Codex, Cursor, and GitHub Copilot are on the list. Gemini CLI, Devin, Factory, Grok, Conductor, Paradev and ChatGPT are also on it.
The code is open source. OpenCrawl says the repository is Apache 2.0, with plugins under MIT. The repo at github.com/getopencrawl/opencrawl-core had 31 stars at the time of writing.
Price and running it
The hosted version is free up to 100 million tokens of session data a month. It does not need a card. Above that limit, OpenCrawl says the price is 0.50 dollars per million tokens. A team can also run OpenCrawl on its own Docker host. The session data stays in that team's own S3 bucket. OpenCrawl says the self-hosted build is free.
The benchmark the company runs
OpenCrawl posts its own benchmark on its home page. It uses the BEAM dataset. BEAM is a set of 2,000 questions. The benchmark scores how well a system answers them from stored history. OpenCrawl says it scored 90.3 percent at 100,000 tokens of history. It says it scored 83.7 percent at 10 million tokens. The same page shows lower figures for unnamed competitors at each size. These are OpenCrawl's own numbers. Treat them as a vendor claim until an independent team runs the same test.
Source
OpenCrawl on Y Combinator Launches. OpenCrawl. OpenCrawl 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.
