Bot Crossing reads your harness sessions and writes one map file
A video game for AI agents. Created by Jarren Rocks
At a glance
- What is it?
- A colony sim that renders every coding agent thread on your machine as a little builder on a hex map, reading each harness's own session files through a per-harness adapter. Seven harnesses are supported, the server runs inside the Vite dev server, and the only file it ever writes is data/colony.json.
- Who is it for?
- Bot Crossing suits people who already run several coding agents on one machine and want to watch them instead of tracking them: it reads real session files, stays on loopback, and persists exactly one file. It is not an agent runtime, since nothing here starts or supervises a thread, and the five harnesses marked not yet are the practical ceiling on coverage.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 22 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 21, 2026, and from our analysis. They are not legal advice.
Star history
- Station-Sciences/bot-crossing906 ★ · +436
| Repository | 2026-09-11 | 2026-11-21 |
|---|---|---|
| Station-Sciences/bot-crossing | 470 | 906 |
436 Stars in the last 71 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
It reads your session files and writes exactly one file
Bot Crossing describes itself as a video game for AI agents, and its central claim is about files rather than features. It reads each harness's own local session files, on your own machine. Nothing is uploaded, there is no account, and it never writes to a harness at all. The single file it writes anywhere is data/colony.json, where the map lives, which makes the persistence story small enough to audit by hand. The dev server binds to 127.0.0.1 by default and answers only its own page, so the thing reading your agent transcripts is reachable from the loopback interface and nowhere else. The root of the repository also carries TRADEMARKS.md and DECISIONS.md, the two files a project adds when it wants its naming and its past calls on the record. The project has 756 stars and 239 forks with 27 open issues, and it has published no GitHub release at all, so there is no tagged build to pin: package.json sits at version 1.0.0 and is marked private.
npm run dev is the whole server
Running it is one command:
npm install && npm run devNode 22.13 or newer is the floor, and package.json states it twice, as a sentence in the README and as engines node >=22.13 next to os darwin, win32 and linux. The scripts section is where the architecture shows: dev runs npm run assets --silent && vite, because the API lives inside the Vite dev server and there is no second process to supervise. start runs vite build && node server/serve.mjs for the built version, serve runs node server/serve.mjs on its own, and test runs node --test across the test directory. Runtime dependencies are two, @mdi/js and three; vite and a pair of gltf-transform packages sit in devDependencies. One metadata detail deserves a second look: the repository url in package.json points at github.com/jarrenrocks/bot-crossing, an owner path different from the one this project is published under.
Adding a harness is one file and one line
Support is per-harness, and the line between supported and unsupported is drawn at the adapter rather than at the feature list. Adding one takes a single new file in server/harnesses/, a single line in its index.mjs, and nothing else. The interface is written down in full in server/harnesses/README.md, together with the thread shape, the ground rules and how to find where a given harness keeps its sessions. The README is also explicit about what such a contribution means: if landing it would require editing the scanner or anything under src/, that means the seam needs widening, and the author would rather fix that than have someone work around it. The colony is the union of every installed harness, and a bot carries the name of the harness it belongs to, so a new adapter adds a new species to the map rather than a filter on an existing one. The Status note above all of this sets expectations early: the project is published as-is, built for the author's own use, and carries no promise of maintenance. Issues and pull requests are welcome but may go unanswered, forking is called an entirely reasonable thing to do, and CONTRIBUTING.md sets out what to expect. That framing is the reason the adapter contract is written down in full instead of promised as a roadmap.
Thread links go through the OS opener, with a terminal fallback
Opening a thread is not something the app does itself. Opening a thread, revealing a folder and starting a new session all go through a harness:// deep link handed to the OS opener: open(1) on macOS, xdg-open on Linux, ShellExecute on Windows. Only the half of the program that opens things needed platform work; the scanning half was portable already. Linux carries an extra step because a desktop app often is not installed there: the scheme is checked first, and when nothing answers it, a terminal running the harness's own CLI opens instead. A setting named Open threads in makes that terminal the first choice rather than the fallback, on all three operating systems. So the terminal path is not a hidden Linux workaround; it is a documented choice you can promote anywhere.
Three supported harnesses have no link to open at all
The support table carries its caveats inline. OpenCode is supported through top-level sessions from its own store, and there is no per-thread link to open, so a bot from that harness cannot be clicked back into place. Hermes is supported per pilot profile and lives in the terminal and chat apps, so it has no link either. Kilo Code is supported for top-level sessions, and a thread there opens as its repo folder in VS Code rather than as the thread. Cursor is supported for agent transcripts, with the composer and sidebar threads explicitly not read yet. Codex opens through codex:// and Antigravity CLI through antigravity://, the latter described as the successor to Gemini CLI, which Google stopped serving individual accounts on 18 June 2026. Five entries are marked not yet: Amp from Sourcegraph, Aider, Goose from Block, Qwen Code from Alibaba and Amazon Q Developer CLI.
One tile per seven threads, and height is transcript length
The mapping between the colony and a real thread table is drawn out in full, and it is the part that makes the metaphor readable. One hex zone is one repo, and a larger repo claims more tiles at one per seven threads, grown as a contiguous blob from the middle outward. One bot plus one building is one session. A bot with no building of its own is an errand that session currently has out, which is to say a subagent. How finished a building looks stands for how large its transcript is, on a log scale, so a long session reads as a taller building rather than a fuller one. Scaffolding marks a site somebody is at right now, walking out of the ship is a thread that just appeared, and walking back into the ship is you archiving it. The ground itself is described as gentle but not flat, running from about -0.3 to +0.24 on the Moon, and half again as far on Mars.
The layout is sticky and every zone is rooted in its first tile
The map is deliberately hard to relearn. The previous arrangement is an input to the next one: a repo that still needs the same number of tiles keeps exactly the tiles it had, one that grew keeps them and claims neighbours, and one that shrank gives back whatever it claimed most recently, so growing and shrinking again returns a zone to precisely the shape it started in. Only a repo that has never been placed is placed at all, and it takes the innermost tiles still free. A zone's origin is its root tile rather than the centre of whatever tiles it happens to hold, so gaining one tile does not drag the zone's buildings, bots and name sideways; the new tile simply appears alongside. Because the arrangement is written to data/colony.json, the layout survives a reload, including for a repo whose last thread you archived.
A zone can be carried, and the allocator once refused
A zone can be picked up and moved. Press and hold one, drag it, and drop it on ground that is free; the footprint follows the cursor as rounded tiles, green where it will land and red where it will not. The interesting part is what changed underneath. A zone that other zones were leaning on used to refuse to move at all, because the allocator throws the whole layout away when the colony breaks into islands. Now the drop stands and whatever it cut loose is slid back into contact, each stranded group moving as one body, so an arrangement made by hand is not reshuffled around you. Two moves are still refused: a drop that lands nowhere near the colony, and one that would leave the zone holding a corner of the map on its own. Before that change the arrangement was a pure function of thread counts, so one session appearing anywhere moved the sort order, the order decided the tiles, and a zone you were watching could jump to the far side of the map because a different repo gained a thread.
Editorial conclusion
Bot Crossing suits people who already run several coding agents on one machine and want to watch them instead of tracking them: it reads real session files, stays on loopback, and persists exactly one file. It is not an agent runtime, since nothing here starts or supervises a thread, and the five harnesses marked not yet are the practical ceiling on coverage. Before depending on it, read CONTRIBUTING.md, because the author states plainly that issues and pull requests may go unanswered and that forking is a reasonable answer. Check the harness table against the tools you actually use, and remember that the layout lives in data/colony.json, so losing that file means relearning the map from scratch.
Frequently asked questions
What does Bot Crossing do with my coding agent threads?
It reads each harness's local session files through a small adapter, on your own machine. Nothing is uploaded, there is no account, and it never writes to a harness at all: data/colony.json is the only file it writes anywhere.
What does Bot Crossing need before it will run?
Node 22.13 or newer, then npm install && npm run dev. package.json declares engines node >=22.13 alongside os darwin, win32 and linux, and npm test runs the suite through node --test.
Which coding agents does Bot Crossing support?
Claude Code, Codex, OpenCode, Antigravity CLI, Cursor, Hermes and Kilo Code are marked supported. Amp, Aider, Goose, Qwen Code and Amazon Q Developer CLI are listed as not yet.
How do I open a thread from the Bot Crossing map?
Threads open through a harness:// deep link handed to the OS opener: open(1) on macOS, xdg-open on Linux, ShellExecute on Windows. OpenCode and Hermes have no per-thread link to open at all, and a Kilo Code thread opens as its repo folder in VS Code.
How do I add support for another harness in Bot Crossing?
One new file in server/harnesses/ and one line in its index.mjs. The interface is written down in server/harnesses/README.md; if landing it would require editing the scanner or anything under src/, the seam needs widening first.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/station-sciences-bot-crossing)Star history chart for your README
[](https://hysenlabs.com/projects/station-sciences-bot-crossing)Was this useful?