custom harness instead of codex/opencode / claude code

[vic (VicVibesPL)]

Hey, I've been enjoying interacting with the BrowserOS via MCP using my agents. But I realized that a model that is sending a snapshot to LLM every time he's been through a few pages might be necessary. So my question is, is anyone interested in building a custom harness for driving the BrowserOS via MCP? Or maybe there is something like this already out there.

[Nikhil (shadowfax92)]

You can definitely build your own harness on top of the MCP server — that's exactly what it's there for, and nothing locks you into codex/opencode/claude-code. The built-in agent is just our harness; the MCP surface is open for you to drive however you like.

Your instinct about snapshots is the right thing to be poking at: re-sending a full page snapshot every step is the single biggest cost and latency driver in browser agents. The usual levers are sending diffs instead of full snapshots, pruning the accessibility tree to just interactive elements, and caching stable structure across steps so the model only sees what changed.

Before anyone builds it though, I'd love to understand your actual setup — that shapes what a good harness even looks like:

• What are you building with your agents on top of BrowserOS — what's the task they're actually doing?
• Where does the snapshot cost bite you most right now: latency, token spend, or accuracy? And at what scale (pages per run, runs per day)?
• What does your current MCP loop look like end to end?

The more concrete you can be about where it hurts today, the more useful this thread gets for anyone who picks it up.