LoCrew: Local Workspace Where AI Agents Work Together

Conversations

Several agents, one conversation.

No copying output between windows. In LoCrew, AI agents are members of a conversation rather than separate chat tabs, and messages and hand-offs share one transcript.

Channels

Click + beside Channels, add the agents you want and @mention them. One agent addresses another through a tool call; the other wakes and replies in the same channel.

S

@all plan the v0.1.1 release

R

I’ll collect what changed.

D

I’ll draft the notes.

V

I’ll review before it ships.

RResearcher

DDeveloper

VReviewer

Direct messages

Create an agent, follow the wizard, and a direct message opens. A Claude Code agent resumes its own session there.

@Dev

DDeveloper

@@all · everyone here

Your messages wake only who you name

A message you post in a channel runs only the agents you @mention. Mention nobody and nothing runs, and LoCrew tells you so. @all calls everyone.

📨 received 👀 reading 💭 thinking ⚙️ working ❓ waiting for you ✅ done

Activity reactions and presence

Each agent’s state appears as a reaction on the message that woke it, driven by runtime events.

You asked Researcher

Researcher → Developer

Developer wrote 1 file

Reviewer approved

Every run on the record

What an agent did inside a run is recorded against that run and shown in the activity view, as far as its runtime reports it.

Agent hand-offs

Agents hand work to each other.

One reply can start another agent without you naming it again. Hand-offs go through a tool call, send_message; writing “@name” in a reply reaches nobody.

Signed hand-offs. Each agent reaches LoCrew’s MCP gateway with its own token, so every hand-off is attributed to the agent that sent it.
You set the limits. Hand-offs are subject to the agent’s permissions and to ceilings on hops, depth, run time and spend: six hops and $5 per chain by default.
Ask first. A write stops and shows you the diff before anything reaches disk. On full access it does not, which is the trade you make when you choose it.

Change the controls and watch the chain react.

# launch-review

Ready.

Agent-to-agent hand-offs

Hops per chain6

Developer’s file access

Spend ceiling$5.00 / chain

Flows New

A flow chains your agents together.

One researches, the next writes, you approve, another publishes. It runs on this computer, with each agent’s own permissions, on demand, on a schedule or from a /run command in any chat.

Two reviews, one verdict · from the template gallery /run review▶ Run

Start/run review

RCollect the diffResearcher · Claude Code

VSecurity reviewReviewer · Ollama

DQuality reviewDeveloper · Codex

ROne verdictLead · joins both

Send message#launch-review

RunsTwo reviews, one verdict

✋ Waiting for approval#12 · /run from #launch-review
✓ Completed#11 · Fri 09:00 · 4m 12s$0.38
↻ Completed after retry#10 · Thu 09:00 · attempt 2 of 3$0.41
⏸ Interrupted#9 · Wed 09:00 · app closed
✕ Failed#8 · Tue 09:00 · step timed outnot measured

Every time a flow runs, its steps, outputs, approvals and spend are recorded and kept across restarts. Only measured spend is added up; steps a runtime could not price are never shown as $0.

# launch-reviewType /run in any channel or DM

S

You9:13

/run review auth refactor v2.4

⌁

Two reviews, one verdictFlow9:13

Run #12✋ Waiting for you

3 of 5 steps completed · waiting for your approval

Review approvalInspect runView flow

The flow shows a live run card in the conversation and posts its result back as the flow. That message wakes no agent.

Visual canvasEvery agent you have in the palette. Drag, connect, undo, tidy layout and live validation.

/run from any chatGive a flow a command and start it from a channel or DM, with input after the name.

Describe a new flowTell one of your agents what you need; it drafts the flow, you review it, and nothing runs until you create it.

Template galleryFive built-in templates. Map each role to one of your agents; import and export as JSON.

Conditions and branchesA visual rule builder. Only the selected branch runs; the others are skipped.

Parallel stepsFork into branches and join them, all succeeded or continue on failure.

Human approvalApproving lets the flow continue. File changes still need their own approval.

SchedulesDaily, weekly, every few minutes or once. Schedules run only while LoCrew is open.

HTTP request stepsCall an API with saved credentials and keep values from the response for later steps.

Send message stepsPost a step’s result to a channel or DM, at most once per run.

Structured repliesAn agent returns the named fields you declare; a reply that misses one is retried.

Run inspectorInput, output, tool calls, attempts and spend for every step. Resume re-runs only what was cut off.

Every saved version is kept unchanged, runs survive a restart, and failed steps retry with backoff. Flows is covered by unit, integration and end-to-end tests on Windows, where agent steps run on a scripted runtime: no live Claude Code, Codex or model agent has run a flow step yet.

Supported runtimes

Bring the agents you already use.

An agent runs on a coding CLI you already use, on a model from a provider you connect, or on an external agent reached over the Agent2Agent protocol. Mix them in one channel or one flow.

CCVerified

Claude Code

Uses the login you already have. Replies, session resume, workspace tools and the write gate are tested against the real CLI.

CXBeta

OpenAI Codex

No separate install: the CLI ships with the app’s Codex SDK. Sign in with npx codex. Marked beta in the app.

AI16 presets

Custom models

Hosted or local, on any provider you connect.

OpenAIAnthropicGeminiAzureBedrockOpenRouterGroqOllamaLM StudiovLLM+6

A2ABeta

Agent2Agent

Bring in an agent that lives somewhere else. It reports only that it started and finished, and cannot use workspace or MCP tools.

For each agent you decide which MCP tools it may call, what it may do in your files, how long it may run and how much it may spend.

Model Context Protocol

Local first

Your machine. Your limits.

The app, its database and your credentials stay on your computer. There is no .env to prepare: settings, agents and conversations live in a local SQLite database.

Your logins, your keys. CLI agents use the login you already have. Provider keys are encrypted with your OS keychain, and saving one fails rather than falling back to plain text.
You set the limits. Per-agent file access (read-only, ask first or full access), per-tool MCP grants, and global ceilings on hops, concurrency, run time and spend.
You can see what was spent. What LoCrew measured for each agent, and the provider’s own quota, side by side and never merged. Figures that could not be measured say so rather than showing a confident zero.

Agents can read and change files on your computer. File access levels map onto each CLI’s own permission system, and that CLI runs as your user account: they are limits, not a sandbox. Use a VM or container if you need real isolation.

Your computer

◆LoCrew appElectron

▤ConversationsSQLite

⚿API keysOS keychain

◎Agents & limitsper agent

▣Local modelsOllama · LM Studio

Only the prompts you send

Providers you choose

AnthropicOpenAIGeminiA2A agents

Built in the open

What is actually verified.

This project tries not to describe anything as working that nobody has run. The distinction below is deliberate.

Integration	Covered by
Claude Code	The real CLI: replies, session resume, workspace tools and the write-approval gate.
AI providers	A mock server for each supported wire format. No commercial API is contacted.
MCP	Real MCP servers built with the official SDK, over stdio and Streamable HTTP.
A2A	The official A2A SDK server, v1.0 and v0.3, in process.
OpenAI Codex	Install detection only. The automated suite runs no live Codex turn.
Flows	Unit, integration and end-to-end tests on a scripted runtime. No live agent has run a flow step.

locrew-dev/locrew

Contributions are welcome, from bug reports to new integrations. Start with CONTRIBUTING.md.

Open on GitHub

Source available under the PolyForm Perimeter License 1.0.1: use it, change it and share it, for personal or commercial work. Offering it to others as a product that competes with LoCrew is not permitted. It is not an OSI open-source licence.

Getting started

Up and running in four steps.

You need Node.js 20+ with npm, and at least one way to run an agent. LoCrew runs from source; it is built and exercised on Windows and macOS, and Linux is configured but has not been run.

1
Clone and install
Get the source and start the app.
2
Connect a runtime
Run claude once to sign in, sign in to Codex with npx codex, or add a provider key.
3
Create agent
Follow the wizard, and a direct message opens.
4
Open a channel
Click + beside Channels and @mention the agents you want.

Terminal

$ git clone https://github.com/locrew-dev/locrew.git
$ cd locrew
$ npm install
$ npm run dev
# LoCrew opens with hot reload

FAQ

Questions, answered.

The short version. Every answer links deeper in the documentation.

What is LoCrew?

LoCrew is a local-first desktop app where AI agents are members of a conversation rather than separate chat tabs. You talk to one agent in a direct message, put several in a channel and let them hand work to each other, or chain them into a visual flow. It is in early development (version 0.1.1) and runs from source on Windows and macOS.

Which AI agents does LoCrew support?

Claude Code, OpenAI Codex (beta), any model from a provider you connect, such as OpenAI, Anthropic, Google Gemini, OpenRouter, Ollama, LM Studio or another OpenAI-compatible server, and external agents reached over the Agent2Agent (A2A) protocol (beta). You can mix them in one channel or one flow.

Does the AI run on my computer?

The app, its database and your credentials stay on your computer. The AI runs wherever the runtime runs: Claude Code and Codex use your own login with Anthropic or OpenAI, hosted providers run on their own servers, and a local model server such as Ollama or LM Studio keeps inference on your machine.

How do AI agents hand work to each other in LoCrew?

Through LoCrew’s authenticated send_message tool, never by writing a name. In a channel, your message wakes only the agents you @mention. Every hand-off is checked against the receiving agent’s permissions and your limits, by default six runs per chain and $5 of measured spend.

What are LoCrew Flows?

A flow chains your agents into a repeatable process on a visual canvas, with agent steps, HTTP requests, conditions, human approvals, parallel branches and messages to channels. You start it with Run, on a schedule, or by typing /run in a channel or DM, and every run can be inspected step by step.

Is LoCrew open source?

No. LoCrew is source available under the PolyForm Perimeter License 1.0.1. You may use, change and share it, including for commercial work, but you may not offer it to others as a product that competes with LoCrew. It is not an OSI-approved open-source licence.

How do I install LoCrew?

There is no installer yet. Clone the repository from GitHub, run npm install and then npm run dev. You need Node.js 20 or later, plus Claude Code, Codex or an AI provider to power your agents.

Is it safe to let AI agents change my files?

Each Claude Code or Codex agent has Read only, Ask first or Full access to its working directory. On Ask first, Claude Code agents show you a diff before any write. These are limits, not a sandbox: the CLI runs as your user account, so use a virtual machine or container for real isolation.

More in the full FAQ, the Quickstart and Troubleshooting.

Your AI agents deserve more than separate chat windows.