Gpu

What gpu MCP server does

The gpu MCP server listing points to a repository whose documented product is branded i, or multiplayer intelligence. The project explores whether many individually controlled intelligences can collaborate through shared questions, local knowledge, evidence, and experiments. Its current implementation is a public laboratory and task-and-evidence ledger, not a documented general-purpose MCP service.

Participants can open questions, take answer-oriented tasks to an AI they already use, preserve the resulting work, and return a linked trace. Public records can be read without an account. The project also describes a later path toward user-controlled connectors, local memory, consented identity, and multiplayer experiments, but those later stages are not presented as completed functionality.

How it works

The repository organizes work into questions, contributions, journals, hypotheses, experiments, results, and follow-up questions. A question can begin as a public task pack containing the exact question, context, requested next action, provenance, and related work. Evidence moves through the lab as participants record observations and other people independently verify or challenge them.

The public data layer provides records through JSON and JSONL feeds, plus a corpus JSON document combining public questions, traces, and events. An append-only event map preserves chronology while typed links connect new traces to earlier public work. Private moderation data, withdrawn records, and pending material are excluded from those public feeds.

The gpu MCP server documentation also describes a synthetic swarm experiment and a physical-device experiment. The physical experiment uses a private three-slot room and local browser or headless Python nodes to train small local weights; it is explicitly not a language-model or H0001 experiment.

Setup and configuration

The documented install path is for the Pocket i Lab Codex plugin, not for an MCP package. Add the repository marketplace and plugin with:

codex plugin marketplace add yukakust/joinmultiplayer.ai --ref agent/game-loop-v0.1
codex plugin add pocket-i-lab@joinmultiplayer-lab

After reviewing and trusting the lifecycle hook, start an E002 run from a new Codex task with $pocket-i-lab start E002 as <pseudonym>. Use $pocket-i-lab finish to close the public run. Participation requires explicit per-task consent. Inactive hooks make no network request.

No environment variables or MCP transport configuration are specified in the supplied material. The README does not document installation for Claude Desktop, Cursor, or another MCP client.

Tools and capabilities

The repository documents these capabilities:

• Public question and evidence records
• JSON, JSONL, and corpus feeds for machine-readable public data
• Append-only event mapping with typed links between traces
• Portable task-ready records for use with an existing AI
• Codex lifecycle journaling through the Pocket i Lab plugin
• Public experiment journals with redacted prompts, answers, statuses, and relative changed filenames

The public journal excludes raw reasoning, tool arguments and results, command output, file contents, environment data, absolute local paths, session identifiers, credentials, and the private publication key.

Limitations and notes

The gpu MCP server is not shown in the README as exposing named MCP tools, an MCP endpoint, or a standard stdio or HTTP transport. The repository instead documents a laboratory workflow and a Codex plugin. Do not assume that opening a task pack downloads an agent or connects to a live network.

The current deployed question flow supports answer-type tasks. Other contribution types remain undisclosed until their return formats are available. The project emphasizes local ownership, inspectability, human approval, and public-by-default contributions; contributors are directed to review the ethics and privacy documents before sharing data.

Code is licensed under MIT, while original documentation and laboratory data use separate terms in DATA_LICENSE.md.