Grove is a Context Harness for every coding agent.
Ten agents can work in parallel, each in its own worktree. You steer them by capturing, organizing and injecting the context they work from.
Website · Quick start · Install · Capabilities · Plugin guide
The idea
You no longer write the final artifact. The agents do. Your job is to manage what they know.
ACP standardized how software talks to AI. Grove is the layer above it: it replaces the client with context, so the triad becomes human ⇄ context ⇄ agent. Every feature in Grove is one move on that context.
| Move | What it means | Where it lives in Grove |
|---|
| Capture | Outside context comes in, brought by you or fetched by the agent. | Studio and Shared Assets, sketches, web search, Figma and any MCP fetch |
| Organize | Notes are curated into memory that lasts and gets reused. | Project Memory: agents leave notes, an organizer merges them into linked memories |
| Inject | Context reaches the agent. You push it, or the agent pulls it. | The composer (files, sketches, memories), Agent context, Grove's MCP tools |
| Negotiate | Review comments, surveys and permission requests settle the result. | Code Review, the Dashboard's Needs you list, commit → rebase → merge → archive |
Running ten agents at once is how you lay the work out. Shipping fast is what you get when you feed them good context.
A tour
Screenshots below come from a demo project and switch with your GitHub light or dark theme. Two of them, the Agent Graph and Buddy, are illustrations from the website, because they need live agent sessions.
Dashboard: what needs you
Start each day here. Needs you lists the permission requests and surveys agents are waiting on, and you can answer them inline. In flight shows every task with its branch and diff. Agent context shows exactly what each new session starts with: workspace instructions, Memory, MCP servers and skills, with their share of the token budget.
Every coding agent, in parallel
Ten agents are built in: Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, Junie, Trae, Kimi, Qwen and OpenCode. Hermes, Kiro and OpenClaw are ACP-ready, and anything that speaks ACP over stdio or HTTP plugs in with one entry. Each task gets its own git worktree and terminal session, so agents don't collide. Custom Agents are reusable personas: a base agent, a system prompt and an effort level.
When agents need each other, they delegate through the Agent Graph. A planner spawns a coder, the coder spawns a reviewer, and replies route back as typed <grove-meta> messages, not strings glued through prompts. Cycles are caught at spawn and only one message is in flight per edge. The graph is exposed as MCP tools, so an orchestrator agent can drive it in code.
Project Memory: context that compounds
Chat history records what happened. Memory keeps what should matter next. Agents leave short notes as they work: decisions, corrections, lessons, each tied to the task and chat it came from. On a schedule, or early once enough notes pile up, an organizer agent merges them into durable memories that are titled, tagged and linked. New sessions search Memory before they start, and the composer shows exactly which memories were read this turn.
Expand the mind map and the same memories become a sky you can turn, zoom and pin.
Buddy: a teammate that stays on
A Buddy is a long-lived assistant with a name, a role and its own Memory, in one continuous conversation. It does small things itself and hands session-sized work to task agents, then follows that work back to you. It runs routines on a schedule, and you can bind it to Feishu, Lark or WeChat to talk to it from there.
Studio: for the work that isn't code
Studio is a project type for briefs, brand guides, specs and layouts, with no git worktree. Sketches are Excalidraw canvases with checkpoints that agents can read and build against. Shared Assets are hard-linked into every dependent task, so each agent reads the same files, and Instructions tell every agent how to work here.
Code Review: settle the work in context
Line-level threaded comments on each task's diff, resolvable by you or the agent. Select a batch of comments and send them to the agent: one diff addresses all of them, and you approve that diff instead of each comment. Commit, rebase, merge and archive follow in one step, with squash-merge detection.
Extensions and plugins
Install skills and MCP servers once, then choose which agents get them, per project or globally. Plugins open the policy layer: a plugin is a folder with a plugin.json, built against a typed SDK. It can add a panel or sidebar page, an MCP server, skills, or a backend that subscribes to Grove events. See the plugin guide.
Five surfaces, one binary
The same tasks, agents and worktrees everywhere:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
