← All writing

Technology

OpenCrawl Gives Coding Agents a Shared Memory

Mani G 3 min read

Four small robots in teal, copper, slate, and cream gather around a workbench, connected by thin glowing threads to a central bowl of warm light.

Picture a developer who spends a decent slice of the week as a switchboard operator for machines. One coding agent is on their laptop, poking at an auth bug. Another is in a cloud sandbox, poking at the same auth bug. Neither knows the other exists. The developer is the only one who does, so they copy-paste summaries between windows like a slightly tired intern. If that sounds familiar, you already understand the problem OpenCrawl is trying to fix.

OpenCrawl launched on 16 November 2026. It is an open-source, self-hostable tool from a two-person Y Combinator company in San Francisco, founded by Shalin Patel and Deep Utkarsh. Their pitch is short: multiplayer for your coding agents.

What OpenCrawl actually does

The idea is simple. Today, an agent's context lives in one session, on one machine, inside one tool. Claude Code keeps its history. Codex keeps its own. Cursor keeps a third. Your teammate's laptop is a fourth walled garden. People become the bridge. The more agents you run, the more of your day goes into carrying context from one place to another.

OpenCrawl records the full transcript of every connected session, along with tools you hook up such as GitHub, Linear, Slack, and Notion. Other agents can then search that history word for word, not a compressed note that forgot the pull request number. They can also post to a shared message board, flag overlapping work, and agree who handles what. The company says this works across laptops and cloud sandboxes, and across agents including Claude Code, Codex, and Cursor.

A quick example

Here is the concrete picture they use on their site, and it is a good one. Suppose your agent is about to add a sandbox so generated code does not run on your servers. It leans toward Modal. Then it searches OpenCrawl and finds that a teammate already compared Modal and Daytona, picked Daytona, opened a pull request with a client, and recorded a Linear ticket: sandboxes get no internet and stop after ten minutes. Your agent builds on that work instead of opening a conflicting change. You find out at review time for a happier reason than a merge conflict.

I like this more than I expected. I am usually allergic to a new layer in the stack. Memory tools that summarise a session into a few bullet points throw away the parts you actually need later: which option was rejected, why, and where the code landed. Keeping the raw session is a boring design choice. It is also the right one.

There are catches. A shared history can get noisy. The day it launched, one developer asked a fair question: who cleans that history? Session logs also hold source code and internal decisions, which is why the self-hosted path matters. You can run it in Docker on your own servers, with data in your own S3 bucket. If you would rather not operate it, OpenCrawl Cloud starts free, with 100 million tokens of session history ingested each month and no credit card required. After that, the official site lists $0.50 per million tokens.

I have not run it in production yet, and I will not treat a vendor benchmark as settled science. The official site claims a lead on BEAM, a memory test over histories from 100 thousand to 10 million tokens, while noting that models and judges differ across published runs. The more interesting claim sits on their Y Combinator page. They are a team of two. During a GitHub Actions outage, they say their agents were blocked from deploying, coordinated through OpenCrawl, put one agent in charge of integration testing, and asked for permission to deploy by hand. Then they tested and shipped.

That last detail is the part I keep turning over. If your agents can already hold a standup without you, what is your job in the room? Maybe it is still the same job: deciding what is allowed to ship. I would like to hear how you would draw that line.