Foxwarm

It keeps working after you close the laptop.

Foxwarm runs your AI agents on a server you own. One task can run for days, spread across your machines, and report back wherever you are: in the browser, on your phone or in your team’s group chat.

curl -fsSL https://foxwarm.net/install-foxwarm.sh | bash

Linux x64 (including WSL) and macOS arm64. Downloads a prebuilt release and starts it in the background, using tmux when available. Includes .NET when no compatible runtime is installed. Packages and requirements

One Tuesday, one task
  1. laptop

    You ask: CI on main is red. Find out why and fix it.

  2. build-box

    It hands the test run to a child session, which runs the suite on your build server.

  3. You close the laptop and go to a meeting. The work carries on.

  4. phone

    You open the child session on your phone and steer it: Keep full precision in prorate().

  5. team chat

    It posts a two-line summary in your team’s group. A colleague asks a follow-up; it answers from the full history.

Every step is a session you can open, read and talk to. How Foxwarm works →

The Foxwarm WebUI in Conversation view: the user asks an agent to fix a failing test suite, and the agent reports the cause, the fix and the commit. The sidebar lists other sessions, some with child sessions.
The same task in the Foxwarm WebUI. Real screenshot, demo project.
Works with
  • OpenAI-compatible APIs
  • Anthropic
  • Local gateways
  • MCP
  • Telegram
  • Matrix
  • QQ
  • WeChat
  • WeCom

01Long-running sessions

Long conversations stay sharp.

When a session nears the model’s context limit, Foxwarm compacts older turns into layered summaries instead of dropping them. The original messages stay in the archive, and the agent can pull back the exact text when a detail matters again. Sessions also survive restarts.

  • Layered summaries that point back to their sources
  • Exact recall of archived messages
  • Optional semantic search across past sessions
How compaction and recall work →
A Foxwarm session after compaction. Asked for the exact original output behind a support ticket, the agent searches the archive with the recall tool and quotes the line the earlier command printed.

02Brain and hands

One brain. As many hands as you need.

The agent loop, memory, models, Skills and MCP connections live on your Foxwarm server, configured once. Nodes are the hands: a build server, a laptop, a container or a browser tab that receives commands and sends results back. Adding a machine means running a small client — no API keys, Skills or MCP setup on every box.

  • Models, Skills and MCP configured in one place
  • Lightweight Nodes on Linux, macOS, Windows, Docker and the browser
  • Move a session to another machine; the conversation stays put
Set up a Node →

When you need tighter control

Keys stay with the brain
Agents can be configured to use models and MCP tools normally without ever being able to read their credentials.
Rules per agent
Allow or deny tools and Nodes for each agent, on top of the instance-wide authorization rules.
Agents that organize themselves
An agent can create sub-agents and give each one part of its Nodes and permissions. No level can exceed what it was given.
An agent switches its session to the remote Node build-box and runs commands there.

03Sessions and agents

Delegate to a team you can see into.

When an agent hands off a task, the work runs in a real session: it keeps its own history, survives restarts and sits under its parent in the sidebar. Open any of them to see what it is doing, or message it directly to change course. Work can also go to another agent with its own memory, rules and machines.

  • Every delegated task is a full session you can open and steer
  • Hand work to a different agent, not just a copy of the same one
  • Agent and Node trees where each level can only narrow what it was given
Agents, sessions and memory →
A session that starts two child sessions in parallel — one on the build-box Node, one under a separate docs agent — waits for both and summarizes their reports.
Inside a child session: while it works, the user sends it a direct instruction and it adjusts its fix.

04Channels

Work it out here. Explain it in the group.

Any chat can be attached to any session. Work through a problem in the WebUI, add the bot to your team’s group and run /attach: it answers with the full history behind it, and anyone in the group can ask a follow-up. Works with Telegram, Matrix, QQ, WeChat and WeCom.

  • /attach connects a chat to an existing session
  • The agent sees who is asking and keeps the whole context
  • Sender allow lists; open a group to everyone while commands stay restricted
Connect a channel →

Everything an agent needs, in one place.

Terminal and editor
Real terminals run in the browser next to the conversation, with an optional VS Code–based editor.
Agent memory
Each agent keeps long-lived instructions and knowledge separate from any single conversation.
Skills
Write a workflow once as a Skill; the agent loads it only when a task needs it.
Any model
Use OpenAI-compatible or Anthropic endpoints, including local gateways, and pick a model per session.
ToolScript
Let an agent combine many tool calls in one small script instead of many round trips.
Timers
Give a session one-off or recurring timers so it checks in, reports or follows up on its own.

Up and running in a few minutes.

  1. Install

    Run the installer for your platform. It downloads a release, starts Foxwarm and prints a sign-in link for the WebUI.

  2. Add a model

    Open Setup → Models and enter your provider, model ID and API key.

  3. Start a session

    Chat with the default agent, give it a working directory, and add Nodes or Channels whenever you need them.

Program files and data live in separate directories, so upgrades and backups stay simple. Full installation guide →