Buzz Mesh and where prompts go
Mesh pools community members' idle GPUs into shared AI compute. A member turns on sharing, chooses a model suited to their machine, and other members' agents can then use it as an ordinary local provider — no API key, no per-token bill.
The trust boundary is specific and worth stating precisely: prompts are processed on the hardware of community members. The relay routes but never sees the tokens. Membership of the community is the only gate.
Why it matters to you
This is the cheapest way to run the high-volume half of an agent team, and it is a distinct privacy posture rather than simply a better or worse one. No vendor sees your prompts, which is stronger than an API in one respect. People you know process them, which is different in another.
For client work the distinction matters because "no vendor sees this" and "no third party sees this" are not the same statement to make to a client.
How to apply it
Decide per category of work rather than once for the whole workspace.
Use it for summarising, formatting, routing, checking and first drafts — where cost per token matters more than peak reasoning quality.
Keep confidential material off it, unless you have established that the community members whose machines would process it are an acceptable audience.
Take the model suggestion Buzz makes for the machine's specifications; an oversized model on modest hardware produces slow poor answers rather than better ones.
Understand that sharing is symmetrical in spirit: if you use community compute, someone is providing it, and if you provide it, you are processing other people's prompts.
If you ignore this
Adopted thoughtfully, mesh removes the cost floor from routine agent work and the API-key requirement from onboarding a colleague.
Adopted by default, it silently changes who processes confidential material — a decision that was never made explicitly and would be difficult to explain if asked about later.
Examples
A shared workshop. The tools are free to use and the work happens on premises belonging to people you know — fine for most jobs, and a conversation for the confidential one.
A four-agent team where the router, the summariser and the checker run on shared compute, and only the analyst runs on a paid strong model. Most of the token volume costs nothing.
Setting shared compute as the workspace default during setup, then processing a client's health-sector material through it because nobody revisited the choice.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026