Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

12 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

Grove

Grove

The terminal workspace manager for AI coding agents. Spin up a forest of isolated agent workspaces. Reach any of them asynchronously from your terminal, your browser, or another agent.

CI Docs Python 3.12+ License: MIT

Supports Codex and Claude Code.

🌳 Overview

Writing the code stopped being the slow part. Deciding what to build did. More agents would help, but two in one checkout overwrite each other's files and fight over the same branch. Grove gives each an isolated workspace and touches nothing about the agent itself. One worktree, one branch, one window. Add a container and it gets a whole stack of its own, prefabricated from your repo and identical anywhere, so permissions off risks a sandbox and not your machine.

Then work arrives on its own. Assign an issue on Linear, GitHub or Gitea and a workspace starts against it, using that issue as its spec. It reports its phase back on the same ticket, so one glance says what is moving, what needs you and what it cost, without anyone reading code to follow along. Shape the next one. Review what came back. Unblock what stalled. You stop writing the code. You do not stop owning it.

✨ Features

Terminal UI

Grove is a terminal program first. Run grove in a repo and create, attach, pause or kill a workspace in one keypress each, while a peek rail mirrors the selected agent's pane live next to its git position.

Docs β†’

The Grove TUI: a project-scoped workspace list with a live agent peek rail showing the summary, recent commits, and transcript

Containerized agents

Optionally, run agents inside devcontainers. Each gets a complete stack of its own, Docker in Docker, so it starts its own database and services without touching yours. It comes up out of your project's .devcontainer/, under resource and egress ceilings that bound both what an agent can reach and what it can spend.

Docs β†’

Grove creating a containerized workspace: the create modal selects the Container runtime after finding the project's .devcontainer/devcontainer.json, then Claude Code runs inside the container behind a DEV CONTAINER statusline reporting the container user, the branch, and the container's own CPU and memory limits

Drive Grove from any agent

Grove ships an MCP server. Claude Code, Codex, OpenCode and other orchestrators create workspaces, dispatch tasks, send follow-ups and steer the fleet β€” your agents managing work at the scale of a whole project.

Docs β†’

Claude Code listing Grove's MCP tools: create, list, peek, pause, and steer workspaces

The whole fleet in your browser

Reach every agent over your network to read a transcript, steer a run mid-flight, or spin up a workspace from the fleet grid, all behind a one-time device pairing that keeps the daemon on loopback.

Docs β†’

Grove's web dashboard alternating between two views: a flat, attention-sorted grid of workspace cards beside a session rail, and a single workspace showing its agent transcript next to the work panel

Status lands on the ticket

Every ticket a Grove workspace works gets one comment, rewritten in place as the work moves. The phase, the checklist and the latest commit stay current where your team already looks. Nobody has to ask, and there is no second dashboard to check.

Docs β†’

Grove's status comment on a tracker issue, showing a table of phase, checklist, branch and commit above a six step progress diagram running from Scoping to Done

Audit and tune the whole fleet

Agents now act on staging and production unattended, and off-the-shelf harnesses leave no record you can hold them to. Grove turns each turn into one trace, so a team can see what ran where, what it cost, and which tool calls are eating the time. Point it at Langfuse or any OTLP backend.

Docs β†’

One Claude Code turn as a Langfuse trace: an agent-turn root with cost and duration, nested model generations, and a span for each Bash, Skill, WebFetch and WebSearch call, with GenAI and Grove attributes on the selected span

Also in the box:

  • Branch-aware lifecycle. Create, pause, resume, respawn, and kill. Pause keeps the branch and drops the worktree. Kill deletes only branches Grove created, never remotes.
  • Configuration cascade. A committed .grove/config.json sets the team baseline. Six layers let each developer override locally without touching it.
  • Per-agent model selection. Each agent exposes its own model catalog. Pick the model per workspace: a cheap one for scaffolding, a strong one for the hard refactor, side by side.
  • Session recovery. Grove pane-verifies each agent session and re-adopts it across daemon restarts and a different user attaching. A stale pointer remaps to the recovered session in a click.
  • Ticket providers. Attach a GitHub, Gitea or Linear issue or pull request to a workspace, or let Grove read the link straight off the branch name so it follows the branch.
  • Push notifications. Get pinged when an agent finishes a turn or needs you.
  • Subscription windows. Every plan's usage, reset and burn rate in one place, with a forecast of which limit you hit first.
  • Usage trends. A year of tokens by day, cost and latency per model, and where the time goes across agents and projects.
  • Local by default. The audit is projected from transcripts on your own disk. Export is opt-in, and prompt bodies stay off unless you enable them.
2-grove-web.mp4

πŸš€ Get started

Grove needs git and tmux, and installs to your PATH as grove straight from the repo. With uv:

uv tool install "grove[daemon] @ git+https://github.com/bearlike/Grove"

cd path/to/your/repo
grove config init        # scaffold .grove/config.json
grove                    # launch the TUI
uv tool upgrade grove    # update later

No uv? Use pipx install "grove[daemon] @ git+https://github.com/bearlike/Grove" or plain pip install --user. See Get Started for prerequisites and every install path.

πŸ€– Let an AI agent configure Grove for you

Configuration has a few layers and many knobs, so you do not have to write it by hand. Hand the prompt below to Claude Code, Codex or any coding agent. It reads Grove's config skill and sets things up with you, then verifies every field against the version you actually installed.

Read https://raw.githubusercontent.com/bearlike/Grove/current/src/grove/skills/configuring-grove/SKILL.md

It is the skill for configuring Grove, a terminal workspace manager for AI
coding agents. It covers the six-layer cascade, agents, init scripts,
containers and ticket providers.

Help me write my Grove user and project config, then verify every field
against my installed version with `grove config schema --stdout`.

πŸ€– Skills for Claude Code

claude plugin marketplace add https://github.com/bearlike/Grove.git
Teach your agent to drive Grove and to configure it for a project

Four skills, shipped from the repository that implements them, so they never drift from the code. They are the same files grove skills install copies into a tool's skills directory.

Skill For
using-grove Driving a fleet from outside. CLI verbs, MCP tool ids, ticket and PR linking
working-in-grove The agent inside a workspace, reporting phase and keeping attached tickets current
configuring-grove The config cascade, agents, init scripts, containers, ticket providers
reinstalling-grove When an update did not take

πŸ“š Documentation

Full documentation lives at https://bearlike.github.io/Grove/latest/.

Section Covers
Get Started Install, prerequisites, first run, verify.
Configure Project setup, agents, init scripts, configuration reference.
Use TUI tour, CLI, web IDE, authentication, daily workflow.
Capabilities Lifecycle, branch provenance, live activity, status semantics, configuration cascade.
Develop Architecture, public API, engineering principles, contributing, design system.
Troubleshooting Symptom β†’ cause β†’ fix for common failures.

🀝 Contributing

Bugs and feature requests on the issue tracker. For development setup, lint/test commands, and PR conventions, see the contributing guide and CLAUDE.md.

πŸ“„ License

MIT Β© Krishnakanth Alagiri.

About

πŸ€ Opinionated Agents Development Environment for Teams. Run a forest of agents in parallel, manage, and monitor them.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Contributors

Languages