innernet

innernet

the memory your AI tools share. innernet keeps one versioned memory for every project you work on and for you — what you're building, what you decided and why, what's in motion — and every tool you connect (Claude, ChatGPT, Codex, Cursor, any MCP client) reads and writes that same memory. open a project in Claude and the conversation picks up where your other tools left off.

this plugin brings that memory into your AI tool: the innernet connector plus two skills that tell the assistant when and how to use it.

what's inside

• innernet connector — the remote MCP server at https://innernet.live/api/mcp. you sign in once through innernet's own OAuth page in your browser.
• innernet skill — find the project you're talking about, load it, answer from the part of the memory that matters, save decisions back, keep the project's task list current, read the rooms you share with other people, move your memory in from this assistant or export all of it, and remember durable things about you that you choose to share.
• innernet-handoff skill — "where did I leave off?" at the start of a session, and a short handoff note at the end, so the next tool continues.

use it

one repo, every host. the same connector and the same two skills, packaged for each:

claude.ai · cowork — customize → plugins → add, then connect innernet from the plugin's connectors tab and sign in.

claude code

/plugin marketplace add your-innernet/innernet-mcp

/plugin install innernet@innernet

codex

codex plugin marketplace add your-innernet/innernet-mcp

codex plugin add innernet@innernet

cursor — install innernet from the cursor marketplace, or add this repo as a plugin.

grok bot — grok bot installs plugins from the cursor marketplace, so the same listing is one press there. without it, ask your bot: *add the custom MCP server innernet at https://innernet.live/api/mcp*, press connect on the card it shows, and sign in.

grok · perplexity · le chat — add a custom connector (grok.com/connectors, perplexity's plugins, le chat's connectors) and paste https://innernet.live/api/mcp.

hermes agent — hermes plugins install innernet (the hermes plugin catalog), then add the server to ~/.hermes/config.yaml with auth: oauth and run hermes mcp login innernet (see hermes/README.md).

openclaw

openclaw mcp set innernet '{"url":"https://innernet.live/api/mcp","transport":"streamable-http","auth":"oauth"}'

openclaw mcp login innernet

gemini cli

gemini extensions install https://github.com/your-innernet/innernet-mcp

any other MCP client — add the remote server https://innernet.live/api/mcp. it is listed in the official MCP registry as live.innernet/innernet.

every app, with the shortest way into each: https://innernet.live/docs/everywhere

you need a free account at innernet.live. never paste an api key into a chat — the sign-in handles it.

try:

• "load my innernet project context"
• "what did we decide about pricing? check innernet"
• "where did I leave off?"
• "hand off — save my progress"

what's in the repo

data

• the plugin runs nothing on your machine: no scripts, hooks, or local servers. it is two markdown skills and one connector entry.
• the only destination is innernet.live. Claude sends the connector what a tool call carries — a project name, your question, or the text you ask it to save — and innernet returns what your memory holds. innernet never reads Claude's own memory, chat history, or files. when you ask to bring your memory over, Claude sends what it already knows about you, and you see what moved in.
• inside innernet, a few model providers do work on what you saved: Anthropic writes and folds, DeepInfra makes the embeddings retrieval uses, and TypeSafe's decision model, reached through OpenRouter with zero data retention, picks which saved sections answer an ask and sorts captures. the privacy policy names each one and what it sees.
• saves are written to your innernet account, versioned, and visible on your dashboard; you can edit or delete any of it there. personal memory is gated by disclosure levels you set, and private facts are never returned to an AI.
• tools that change data ask before running, and the skills tell Claude to say what it saved.

privacy policy — https://innernet.live/privacy · terms — https://innernet.live/terms

releasing

every push to main is a new version for the plugin directories, so each one raises version in all five plugin manifests (claude, codex, cursor, gemini, hermes) together. the hermes catalog pins an exact commit: after a release, bump sha in plugin-catalog/innernet.yaml on NousResearch/hermes-agent with a new PR. python3 scripts/check.py runs on every push: the manifests parse, the versions agree, every url is https://innernet.live/api/mcp, the listing fits each directory's caps, and the skills name only tools the live server serves.

server.json is the official MCP registry entry. its version follows the server's own serverInfo.version; when that moves, raise it here and run ./scripts/publish-registry.sh (it signs in with the innernet.live domain key).

support

docs — https://innernet.live/docs/connect · email — yours@innernet.live

MIT licensed.