Share and use shared compute
Everyone · Automate the work
Steps at a glance
1. Understand the trade before you opt in
Mesh pools community members' idle GPUs into shared AI compute. Agents see it as an ordinary local provider, so nobody needs an API key and there is no per-token bill.
The trade is where the prompts go: they run on other members' hardware. The relay never sees the tokens, and membership is the only gate. That is a meaningfully different trust model from a vendor API — better in that no company sees your prompts, different in that people in your community's machines process them.
For client work, this is a decision, not a default. See Handle sensitive data where agents are present.
2. Share your machine
Open Settings, go to Compute, choose the model Buzz suggests for your machine's specifications, and turn on Share this machine.
Buzz's suggestion is based on your hardware; a larger model on an underpowered machine mostly produces slow, poor answers rather than better ones.
Sharing is symmetrical in spirit but not required — you can use community compute without sharing, though a community where nobody shares has no compute.
3. Set the shared provider as your agent default
In the Agents area, set the default LLM provider to the Buzz shared compute option. Agents created afterwards inherit it and need no API key.
4. Point an agent at it and test
Create or reconfigure an agent to use the shared provider, add it to a channel, and give it a small request.
Expect the first response to be slow if the model is still downloading. That is the most common reason a newly configured agent appears dead. See Stop, steer or recover a runaway agent.
5. Decide what should not run there
Shared compute suits high-volume, low-stakes agent work — summarising, formatting, routing, first drafts — where cost per token matters more than peak reasoning quality.
Keep two categories off it: work needing your strongest model, and anything whose content should not be processed on someone else's machine.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026