# GamerAI Community Terms of Service *Plain-English version. Last updated 2026-08-22.* GamerAI is a community-driven AI service. People donate spare GPU time to a shared pool; their invitees and (in time) paid customers use that pool to ask questions of language models. We think AI should be something users can be proud to use and run on their own hardware, not just rented from a handful of hyperscalers — so we're building it that way. These terms apply to every person who interacts with the network: **contributors** (who run the agent on their machine), **invitees** (who use the network through a contributor's invite), and **admins** (who operate the coordinator). Accepting an invite, installing the agent, or operating the coordinator means you agree to these terms. --- ## 1. Honesty, both ways ### As a contributor - **Run the unmodified agent and an unmodified inference engine.** Do not swap in a finetuned model, do not wrap inference calls, do not tamper with the output before returning it. We periodically send canary prompts with known-good answers to detect modified models; contributors who fail canary checks are flagged and may be removed. - **Advertise capabilities truthfully.** The model your agent reports on registration is the model your agent actually serves. Don't claim to run a 13B-class model while serving a 1B one. - **You will see the prompts you serve.** Treat them like you'd treat a stranger's diary you stumbled across. Do not log them outside the agent's own files, do not share them, do not search them. The agent rotates its log file automatically; you can delete it any time. - **Do not enable inference-engine debug logging.** Specifically, do not run Ollama with `OLLAMA_DEBUG=1` (or any equivalent verbose flag) while your agent is serving the community queue. The agent itself does not log prompt text, but debug mode does — and that is the most plausible way prompts could leak from your machine. The agent spawns its inference engine with debug off by default; a contributor who launches it themselves controls its environment. - **Don't game the network.** Don't run multiple agents on the same machine for inflated tier counts. Don't claim work you don't intend to serve. Don't intentionally fail jobs. ### As an invitee or paid customer - **Don't send secrets.** Treat the GamerAI network like a public chat room with friends-of-friends listening: whatever you submit may be visible to whichever contributor's GPU happens to serve it. Do not paste passwords, API keys, PII you wouldn't put on a postcard, or anything subject to NDA. - **Don't abuse the network.** No automated scraping, no illegal-content generation, no harassment, no attempts to exfiltrate prompts from other users. Standard acceptable-use-policy stuff. - **Use higher-confidentiality options when they're available.** We have plans for routing paths that don't require the contributor to see plaintext; until those ship, contributor visibility is what you get. ### As an admin - **Do not log full prompts beyond what's needed for debugging.** Truncate or redact in persisted records. - **Disclose monitoring.** Canary prompts, agreement scoring, and worker reputation are part of how the network stays trustworthy. They are not surveillance of contributors' machines. - **Don't read prompts unless you have to.** "Have to" means debugging a specific incident, not curiosity. --- ## 2. What we monitor, and why The coordinator periodically does three things, and you agree to them by joining: 1. **Canary prompts.** A small number of normal-looking prompts with known correct answers are mixed into the queue. Whichever contributor claims one of these has their response compared to the expected answer. Repeated misses flag the contributor for review. 2. **Agreement sampling.** Occasionally the same prompt is served by two contributors and the outputs are compared. Contributors whose answers consistently diverge from peer consensus are flagged. 3. **Capability auditing.** The coordinator may query the worker's inference engine (via the agent) to confirm the advertised model is the one actually loaded. None of these read into prompts on the contributor's machine. They detect tampering by observing what the network sees. --- ## 3. Service guarantee GamerAI is provided on a best-effort basis. The service may go down, prompts may time out, contributors may sometimes return poor-quality output. We work to minimize all of those, but we make no guarantees about uptime, durability, or correctness. Infrastructure will evolve over time, and reliability may change as it does. If you're using GamerAI for anything where downtime or wrong answers would cost real money or harm real people, **don't.** Use a paid service with a contractual SLA. Paid-tier offerings with stronger guarantees may follow; until then, plan for zero. --- ## 4. Earnings, money, and the paid pool Today, contributors are **not paid** for serving the free community pool. Tier, status, and invite slots are what the free pool rewards. We have plans for an optional paid customer layer where opt-in contributors will earn a share of paid revenue for paid jobs they serve. The exact structure of that is still being designed; any "earnings" figure shown in your agent dashboard today is a usage ledger, not a contractual obligation to pay you anything. It exists so that when paid features ship we can apply them in a backwards-friendly way. We intend to build GamerAI into a sustainable business. The free community pool is the foundation, with paid options layered on top. Both sides will evolve over time. --- ## 5. Termination We can remove a contributor for failing canary checks, abusive behavior, or material breach of these terms, at our discretion and without notice. You can stop participating at any time by closing the agent. The coordinator does not have remote-kill control over your machine; uninstalling the agent uninstalls our presence on your hardware. --- ## 6. Privacy - **Prompts** traverse the contributor network and are visible to whichever contributor's machine serves them. We log job metadata (timestamps, sizes, durations) and retain prompt content long enough to deliver responses and maintain your history. We do not sell, train on, or otherwise reuse prompt content. - **Contributor identity** consists of an email address (for invite chain) and a bearer token hash. We do not collect hostnames, MAC addresses, or hardware fingerprints beyond the capabilities the agent voluntarily reports — which includes your GPU's model name and VRAM amount (e.g. "RTX 4090, 24 GB"), shown on your account's Machines page and to admins operating the network. That's a shared model descriptor, the same string millions of other owners of that card would report, not a unique identifier — it's used to label your machine in the dashboard and is not cross-referenced against anything to re-identify you. - **Machine naming.** At setup the agent asks for an optional name for the PC ("Living Room Rig," etc.) so you can tell your machines apart on the dashboard. It's entirely your choice what (if anything) you type there — leave it blank and the agent picks a random two-word name instead, nothing about your machine or identity. Whatever you enter is stored and shown back to you (and to admins) verbatim, so don't put anything in it you wouldn't want visible in the dashboard. - **Logs** rotate. Stored prompts are eligible for purge after a reasonable retention window. If you want yours purged sooner, email the admin. --- ## 7. Disputes If a dispute arises, both parties agree to: 1. **Talk first.** Email the admin and try to resolve the issue in good faith. Most things end here. 2. **Then arbitration.** If direct discussion doesn't resolve it, binding individual arbitration through a recognized arbitration service (such as the AAA or JAMS), with each party bearing their own costs except as determined by the arbitrator. Class arbitration is not permitted; disputes are individual. 3. **Court only as a last resort.** If arbitration is unavailable or unenforceable for a specific matter, jurisdiction is the state of residence of the GamerAI operator. This is not a substitute for legal advice. If you have a lawyer, listen to your lawyer over us. --- ## 8. Changes These terms can change. When they do, the coordinator's `/tos` endpoint will publish the new version with an updated date. Existing members will see a notice on next agent log line (for contributors) or next sign-in (for invitees and paid customers) and may be asked to re-accept before continuing. --- ## TL;DR Be honest. Run what you say you're running. Don't read other people's prompts. Don't put secrets into the network. If you break trust, you're out. If we break trust, leave. GamerAI is community-driven — AI infrastructure you can be proud to use and run on your own hardware, not just rented from hyperscalers. The free contributor pool is the foundation; we plan to layer paid options on top and grow this into a real business while keeping the community-first part intact.