apra-fleet MCP
apra-fleet runs a fleet of AI agents across machines and models through MCP and drives multi-hour workflows
- 106 ★ on the repository
- Last commit
Broad access
Why this level
- Runs commands and launches agents on registered machines, including over SSH
- Issues SSH keys to machines and stores secrets that agents use in their work
- Multi-hour workflows run unattended and may commit, deploy and spend paid model credits
Install
- Works in
Manual install
npm install -g @apralabs/apra-fleetThen `apra-fleet install` and `apra-fleet start`. Needs Node.js 22.16+.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Runs commands and launches agents on registered machines, including over SSH
- Issues SSH keys to machines and stores secrets that agents use in their work
- Multi-hour workflows run unattended and may commit, deploy and spend paid model credits
This is third-party code. Review the repository files before installing.
What it does
apra-fleet is an MCP server and process manager for a fleet of AI agents. Fleet members are real machines running Windows, macOS and Linux, local or reachable over SSH. Registration happens in plain language through the agent: "register a member `build-server` at 192.168.1.10". The remote password is typed in a separate terminal, is used only to set up an SSH key and is then forgotten. Secrets are stored separately, out of the model's sight. Members can be assigned different models: Claude, Codex, Copilot, Antigravity or local models through OpenCode. Routing follows cost tiers (cheap, standard, premium), and review is meant to be done by a different model. `apra-fleet workflow hello-world` starts a trial workflow, and the flagship `fleet-sprint` runs a cycle of plan, develop, review, deploy, integration test and harvest until the goal is met. Workflows are long and resumable: they have supervisors, watchdogs, member reservations and a live dashboard, and records of runs and spend are visible in real time. The Claude Code install adds a `fleet-status` mod: member status in the status line and a `/fleet-status` pane.
Who it is for. Engineers and small teams with several machines of their own and subscriptions to different models, ready to hand agents multi-hour tasks with review and tests.
Good fit when
- You want one workflow to run changes across a build server, a laptop and a GPU box
- One model writes code and another checks it, with review by a different provider
- A task runs for hours and a crashed agent must be detected and the work resumed
- A premium model is for review while mechanical edits go to a cheap one
- A local model handles private data and a cloud model runs the rest of the plan
Not a fit when
- There is a single agent on a single machine, where a fleet adds needless complexity
- Agents must not get SSH access to real machines
- You only need a quick parallel launch of several terminal sessions
- The system has no Node.js 22.16 or newer
Example request
Register two local members doer and reviewer, pair them and run the hello-world workflowMCP
- Transport
- http
- Authentication
- not required
Limitations
- Needs Node.js 22.16 or newer. The npm install does not register an OS service, so restart the server with `apra-fleet start` after a reboot.
- The product coordinates agents on real machines, and result quality depends on the models. The author says the repository is itself developed by a fleet, which gives no guarantee for your code.
- The GitHub repository card shows no detected license, while the README says Apache 2.0.
- Other model providers connect with their own install commands, for example `apra-fleet install --llm codex`.
How to disable. Stop the server and remove the package with `npm uninstall -g @apralabs/apra-fleet`. Remove members and secrets through the agent before uninstalling, and revoke the SSH keys issued to machines by hand.
FAQ
How do I install apra-fleet for Claude Code?
Run `npm install -g @apralabs/apra-fleet`, then `apra-fleet install` and `apra-fleet start`. In Claude Code connect the server with `/mcp`; the default address is `http://localhost:7523/mcp`.
How does apra-fleet handle remote machine passwords?
The password is typed in a separate terminal and not in the chat. It is used once to set up SSH keys and then forgotten. Secrets are stored out of the model's sight.
Which models can be mixed in a fleet?
Claude, Codex, Copilot, Antigravity and local models with an OpenAI-compatible endpoint through OpenCode. The cheap, standard and premium tiers control cost.
What does the fleet-sprint workflow do?
It runs cycles of plan, develop, review, deploy, integration test and harvest until the goal is met or the evidence says stop. It starts with `apra-fleet workflow fleet-sprint --issue ... --members ...`.
Which systems does apra-fleet run on?
macOS, Linux and Windows. A fleet member can be a machine on any of them, local or reachable over SSH.
Discussion
No questions or reviews yet.
Related
MCP serversA self-hosted control plane for running several agents: task dispatch, run review, spend tracking, connected via MCP, CLI or REST
An open-core platform: workflows and actions of connected services are published as MCP tools for Claude, Cursor and Windsurf
Flow-Like MCP
Flow-Like
Flow-Like MCP nodes expose visual workflow functions to an agent over HTTP, with owner-selected tools and access
A Discord MCP for Claude and Cursor: the agent posts and searches messages and manages forums, channels, roles and webhooks through a bot token
Discord MCP by cappyeo: a server with 200+ tools
Discord MCP (cappyeo)
A Discord MCP with over 200 typed tools: the agent handles messages, roles and moderation, builds a server from a plan, and writes default to preview mode
The Apideck MCP server: an agent works with over 200 business systems: accounting, HRIS, CRM and files through one unified API
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| apra-fleet (this card) | yes | not required | 106 | |
| CLI Agent Orchestrator | no | 1,515 | orchestrates CLI agents on one machine, with no SSH device fleet | |
| Claude Squad | no | 8,738 | manages several terminal sessions and runs no resumable workflows |
README in short
The README opens by comparing an agent fleet to container orchestration and explains why devices, models, workflows and secret control matter. It continues with the flagship workflow, a five-minute quick start and links to the `fleet-sprint` guide.