The terminal workspace for AI coding agents.
Open Source on MacOS - Live Now
$GRID CA: 0x6B456E66524aEC1792013eF9DFE87e3F84311ba3codegrid.app macOSJoined May 2026
first-run onboarding is 4 steps: welcome → detect your installed agent CLIs → one click to enable Agent Bus → done. from download to a talking fleet in under two minutes.
3/ it's also the escape hatch: a hot new model drops next month? it has a pane waiting. six agents today. the architecture says: as many as the ecosystem produces, eventually.
2/ the result: privacy-focused inference with model choice, in a first-class pane with status, naming, Agent Bus, everything. philosophically this is the local-first thesis again: you choose the inference provider. the workspace doesn't care.
shipped agent #6 (Venice) recently, and the lesson was humbling: adding an agent type touches 14 files. wrote ourselves a checklist so #7 takes an afternoon. boring process beats heroics.
agent #6 is the wildcard: Venice. CodeGrid auto-installs aider under the hood, so one pane can run basically any model. the fleet has an open slot for whoever you want.
4/ it also benchmarks your fleet on your real workload, not synthetic evals. you'll learn who's good at what fast. worktree isolation (more next week) means parallel agents can't trample each other's changes. safety built in. six agents plus shell makes 7 pane types. your fleet, your rules.
3/ five models attack in parallel with five different theories. first one to repro wins. cancel the rest. this turns debugging from a serial slog into a race you spectate.
parallel fan-out is the most under-used agent pattern: give five agents the same hard bug and take the first correct answer. compute is cheap. your friday evening isn't.
my fleet at 5pm: Claude pondering the architecture like a philosopher. Codex on tweet 47 of a refactor. Gemini quietly on test run #212. the shell pane: still my most-used. we are not the same.
you can mix providers per workspace: Claude on the API design, Codex on the frontend, Gemini on tests. one canvas, three vendors, zero app switching. best model per task is the whole point.
3/ pin a markdown note next to the pipeline with the plan. notes anchor to panes, so context travels with the work. cost optics: burn your premium tokens on judgment-heavy stages, let free tiers grind the mechanical ones. the canvas isn't decoration here. spatial arrangement is the pipeline diagram.
2/ concrete version: Claude implements a feature → messages Gemini 'run and fix the tests' → Gemini hands off to Codex 'update the docs'. each stage is visible as a pane. you literally watch work flow left to right across your canvas. kanban, but alive.
6K Followers 764 FollowingDEGEN meme full time. Lost my Twitter for the 3rd time.
Just a regular trader sharing info, not financial advice. Always NFA/DYOR
5K Followers 1K Following“If no mistake u’ve made, losing you are. A different game u should play.”@bankrbot is in the empire-building biz. #bitcoin is simple, everything else complex.