# I built an open-source workspace to tame my AI coding agents > Source: > Published: 2026-07-18 06:33:44+00:00 Every AI coding CLI is a snowflake: different flags, different session models, and all of them editing your one working copy. I wanted to run agents in parallel on isolated branches, queue prompts while an agent works, and have a real UI — without handing API keys to yet another app. [AgentGrove](https://github.com/arnabk/agentgrove) is an open-source (MIT) local developer workspace. It launches the agent CLIs already on your machine (Claude Code, opencode, Kimi), translates their event streams into one UI, and scopes every chat to a git worktree. `pnpm install` et al, celestial branch names with a galaxy map of where you've beenNo vendor SDKs. Each provider is a small subprocess adapter (~400 lines of Rust) that launches the CLI with its JSON stream flags and translates events into a common enum (`Token` , `Thinking` , `ToolCall` , `ToolResult` , `Done` , `Error` ). Auth stays with your local CLI — the app never sees a key. Adding a provider means writing one translator and registering it. Two layers keep this honest: a deterministic fake provider for dispatch/queue/resume e2e tests, and real-model Playwright runs against the actual CLIs. The demo videos on the README are recorded by Playwright in an isolated Docker stack (including a seeded Postgres for the DB editor demo), not staged screenshots. ``` git clone https://github.com/arnabk/agentgrove.git cd agentgrove && just dev ``` Demo videos are on the [README](https://github.com/arnabk/agentgrove). Issues and PRs welcome — I'd love to hear which provider you want next.