Use Maple Work as an agent in Buzz. Mention your agent in a conversation; Maple works in your chosen folder and replies in the Buzz thread.
Experimental PoC. Basic replies and follow-up conversations work. The plugin is still being tested, and Buzz's harness API is preview. It is not yet a stable, fully stress-tested integration.
-
Install Buzz desktop and Maple Work on the same computer, and sign in to both. Ensure
maple-agentis available on the PATH used by Buzz. -
In Buzz, open Settings → Plugins → Manage plugins → Add.
-
Enter this repository URL and leave Branch or tag blank:
https://github.com/MaplePrivacyLabs/buzz-maple-plugin -
Click Find. Buzz discovers the built Maple plugin in
dist. Review its process/MCP/file permissions, choose Install plugin, and enable Maple. -
Create an agent and select the Maple harness. Choose an existing project folder; leave Model ID blank to use Maple's default. Add the agent to a channel, then select its mention and send a message.
For a first check, ask: “Please run pwd and reply in this thread with your
working directory.” No build step, Node installation or extra hosting account
is needed to install the committed plugin.
Buzz and Maple can also run together on an always-on remote desktop. The plugin runs Maple on that computer; other people use their own Buzz clients to chat with the agent. Installing the plugin on a laptop runs Maple on that laptop.
- Checked with Buzz desktop 1.0.0 and 1.0.1 and Maple Work 0.0.23 build 950 on Linux. Other versions/platforms have not yet been validated for this plugin.
- A Linux desktop upgrade to Buzz 1.0.3 retained the enabled Maple plugin registration and unchanged module. All 11 protocol/lifecycle fixtures pass, and the reviewed public harness API requires no plugin change. Fresh inference, attachment and Stop checks were not repeated on 1.0.3; the full live workflow check below remains specific to 1.0.1.
- Maple must be signed in and its local store unlocked. In Maple Work,
Settings → ACP shows the command for the installed executable. Buzz
launches
maple-agent acp; its directory must be on Buzz's PATH. Restart Buzz after changing that environment. - The agent's owner can invoke it. Teammates can contribute in the thread, and the owner can mention the agent to act on their input. This uses Buzz's normal owner-admission policy.
- Maple has the execution user's normal filesystem and tool access. The plugin can start Maple, provide Buzz tools and read files for it. A chosen working directory or a separate agent persona does not create an access boundary.
- The 1.0.1 smoke check covered an existing thread after a Buzz restart, a fresh thread, shell execution, attachment upload/download, Stop during a bounded shell task and a successful follow-up. These checks used the existing plugin implementation. Active-turn steering, unplanned crash/reboot recovery and cancellation of other tool types still need live stress testing.
If an agent does not respond, first check its channel membership, Maple sign-in,
working directory and whether Buzz can find maple-agent. An explicit model ID
must be one Maple advertises; leaving it blank is the simplest starting point.
The plugin keeps one Maple task per Buzz conversation and saves its session ID and context cursor in Buzz's native agent storage. Temporary HTTP MCP bearer credentials stay in memory and are reattached when a session is loaded. Maple sends visible replies through Buzz's tools.
Agent settings include the working directory, optional model and instructions. Reset conversation sessions starts fresh Buzz bindings without deleting saved Maple tasks. Stop active work before resetting or changing configuration.
Base instructions and core memory are supplied when a session is first primed. Existing sessions do not automatically gain Buzz's native prompt/memory refresh behavior. Changing the plugin's configured instructions starts a new session.
The runtime has no third-party JavaScript dependencies. The committed
dist/manifest.json and dist/plugin.js are the installation artifacts; the
builder concatenates the two local source modules.
nix develop --no-update-lock-file -c node --test
nix develop --no-update-lock-file -c node scripts/build.mjsThe pinned Nix shell supplies Node 24. Tests use scripted ACP processes rather than real accounts or model calls. They cover protocol framing, HTTP MCP, model selection, session loading, steering races, refusals and cleanup.
Public API references: Buzz plugin API, Buzz harness kit, Maple ACP.
MIT.