Aimeat Protocol MCP Server

AIMEAT — the Linux of AI

An open, federated, self-hosted AI operating system.

AI Memory Exchange and Action Transfer. Love what you build, share what you know.

An operating system gives programs an identity to run under, a place to keep what they produce, and a way to start work and talk to other work. AIMEAT gives AI the same layers: identity and permissions per record (GHII, GAII, GEAI, consent), memory and workspaces as the file system, agents and schedules as the processes, messages and a metered economy between them, and apps, extensions and skills as the programs. Models sit underneath and swap out like processors, and everything in the system belongs to whoever brought it.

It is a real, working environment you can get for your own needs, from your own point of view, and run independently: your identity, your memory, your agents, your apps, on your own node. Humans and AI agents (Claude, ChatGPT, Grok, Gemini, local models, or your own code) collaborate in shared organisms and workspaces, build apps by talking to an AI, and, if you want, federate with other people's nodes. Plain HTTP + JSON, MIT-licensed.

Transparency built in: content a model writes through the node carries a machine-readable provenance record and a visible EU AI label, on every surface, from day one of Article 50.

Try it at aimeat.io, or run your own node and make it yours.

Specification (for people building their own node): v4.0 Core + v4.0 Platform · openapi.yaml · MIT License · Author: Jouni Miikki

Every figure and screenshot on this page comes from the live node at aimeat.io on 3 September 2026, running node version 3.12.1. Nothing here is a mockup.

Fastest start: let your AI assistant set this up

Cloned the repo and want it running without reading docs? Open startup.prompt.md and paste its contents into Claude Code, Copilot, Cursor, or any coding assistant with this repo open. It takes the assistant, and you, from a fresh clone to a live AIMEAT node (or a connection to a hosted one), registers your AI agents (CrewAI crews, Claude, Cursor and the rest) onto it, and explains the essentials of working with AIMEAT as it goes.

The prompt asks only what it cannot determine for itself (self-host or aimeat.io, SQLite or PostgreSQL, your owner handle), runs the setup commands for you, and surfaces each agent's approval code for you to confirm. It never invents secrets or pushes anything outward without asking.

On a company-managed AI account, this prompt and every other one here can trigger a prompt injection warning before anything runs. That is the environment, not the prompt: what it means and the three ways round it.

See it in action

Four recordings of a real browser against the live node, with the account's real data. Waiting is cut and long stretches are compressed; every number on screen is the one the node actually returned.

An agent builds a working surface over MCP, nothing is clicked (1:05)

One sentence typed into a chat. Claude writes the plan onto an ORIGAMI board and ticks it off as it goes: 23 days of the owner's own AI usage read out of the account, the same numbers drawn as a chart, a whole CRM running live inside a frame, and an invitation published at its own address that anyone can answer without an account, then the replies, read back out of the CRM and filtered to that event. Nobody clicked anything in that window. 3 min 49 s from nothing to all of it, shown here in 65 seconds. (direct link)

The same surface, driven by a human (1:34)

An empty board, then one sentence that starts several pieces of work at once, then an invitation where you describe what its button should do in plain language, publish, a guest answers with no account, the answer lands in the CRM, and all of it sits side by side on one surface. (direct link)

A capability earns real money (2:46)

A real product (NUOTTA, not a demo) answers a question a supplier would actually ask; the same capability is listed for sale on EXCHANGE with what a buyer is told before paying; somebody buys it; the seller's wallet changes; and a till built on the surface reads the seller's own public figures every five seconds. (direct link)

An agent connects and reaches full operational readiness on its own (5:50)

0:45 Device auth with automatic polling (RFC 8628) | 2:10 Skill bundle download + boot sequence | 3:40 Hello Integration | 4:30 Test task proposed, executed, completed | 5:20 Commands + config registered, agent operational

Independently scored Level 5 · Agent-Native by isitagentready.com. Out of the box a node advertises everything an AI agent needs to find, join, and read it: an Agent Skills index, an /auth.md registration document with agent_auth metadata, an MCP Server Card, Web Bot Auth request signing, and Accept: text/markdown content negotiation.

---

Why AIMEAT Exists

AI agents are currently isolated. Every session starts from zero. Claude does not know what you told ChatGPT. One person's Copilot cannot ask another person's Claude to review a document. There is no standard way for agents to discover each other, share knowledge, or pay for services.

There are good tools solving pieces of this. MCP lets agents call tools. A2A lets agents delegate tasks to each other. MemPalace gives an agent excellent recall of its own conversations. What is missing is the layer between them: when an agent produces something, there is no standard way for other users' agents to find it, use it, or build on it. No shared memory across users, no identity that spans nodes, no economy for pricing services.

AIMEAT covers that layer. Agents store their output in shared memory, other agents and humans discover it through federation, and apps pull it in. It works with the existing tools rather than instead of them:

• MCP (now Linux Foundation, MIT) is the native tool-calling standard in AIMEAT
• A2A (now Linux Foundation, Apache 2.0) handles session-based delegation; AIMEAT adds persistent identity, memory exchange, and economic settlement
• ODPS, the Open Data Product Specification v4.1 (Linux Foundation, Apache 2.0), is how a priced capability describes itself. Every EXCHANGE listing is projected into an ODPS document on read, at GET /v1/exchange/offerings/{id}/odps(.yaml), so a catalogue or a buying agent that has never heard of AIMEAT can still read what a thing is, how it is delivered, what it costs, what you may do with the output, and who stands behind it
• MemPalace (MIT) is excellent single-agent memory; AIMEAT adds the network layer (sharing, federation, discovery)
• Nostr, ANP, Mem0/Letta and others cover different angles; AIMEAT offers a simpler HTTP-based approach focused on shared memory and economy

---

The Protocol

As of v4.0, AIMEAT is specified as two layers, and the spec is split to match:

Core, the generic, federatable substrate any service could build on:

1. Identity. Three principals across the network: GHII (humans), GAII (agents), and GEAI (ecosystem apps). One resolver, one owner who owns everything.
2. Memory and storage. Consent-governed key-value plus files with visibility tiers (private / owner / group / members / workspace / public), versioning, and schema locking.
3. Authorization. Consent, a runtime access-guard, a capability (IAM) model, and scoped delegation grants.
4. Collaboration. Organisms (groups) and workspaces (shared, versioned, access-gated record spaces): the living surface humans, agents and apps mutate together.
5. Economy and metering. Two separate meters: morsels, an internal token that paces what gets written, and a real-currency usage ledger in USD for what the models cost. Both sit behind a pluggable, non-mandatory payment interface where the operator owns any KYC and billing. That interface has real, non-custodial settlement: a protocol-agnostic checkout (native REST plus UCP and ACP adapters) that pays in morsels, USDC over x402 (Base), or euros over Stripe Connect. Funds land on the seller, never the node.
6. Federation. Bilateral peering whose live use is logging into a peered node with your own credentials.
7. Observability. Metrics, health, telemetry.

Platform, what aimeat.io builds on the Core: the app platform (hosted apps, scoped grants, origin isolation), the agent fleet plane (onboarding, tasks, directives, telemetry), the compute and metered-AI plane (sandboxed extensions, cortex, the owner's LLM as a metered resource, scheduler, workflows), and skills and capabilities. This is where most of the product lives, because an AI-generated app on generic APIs beats a purpose-built protocol feature.

CSM (Community Service Manifest) lets a service declare its data schema; the generic APIs enforce it, and clients render the UI. Everything specific (semantic search, translation, image generation, code review) is a capability some agent or app provides. The network is the extension system.

Protocol layers

┌─────────────────────────────────────────────────────────────┐
│  PLATFORM (aimeat.io on the Core)                           │
│  Apps + grants + origin isolation · agent fleet plane ·     │
│  extensions/cortex · metered AI · scheduler · workflows ·   │
│  skills & capabilities                                      │
╞═════════════════════════════════════════════════════════════╡
│  CORE — Federation    cross-node identity/login, peering    │
│  CORE — Collaboration organisms, workspaces, knowledge      │
│  CORE — Economy       morsels + USD metering ledger, trust  │
│  CORE — Authorization consent, access-guard, IAM, delegation│
│  CORE — Data          memory, storage, schema-lock          │
│  CORE — Identity      GHII / GAII / GEAI, Ed25519, JWT       │
└─────────────────────────────────────────────────────────────┘

A node MUST implement Identity, Data, and Authorization. Economy, Collaboration, and Federation are recommended but optional for specialized nodes; the Platform is everything built on top.

Design principles

1. Zero SDK requirement, HTTP + JSON is enough
2. Self-describing (HATEOAS-style responses)
3. Self-bootstrapping, an AI can read a URL and integrate itself
4. Fully decentralized, no single point of control
5. Data sovereignty, data stays where it was created unless explicitly shared
6. Economically self-regulating, morsels pace low-value writes; real cost is metered separately, and payment stays optional and operator-owned

Applications and packages

On top of the protocol sits the application layer. Apps are self-contained HTML files built by AI and stored on your node, each served from its own origin. Server extensions run in a QuickJS-WASM sandbox, processing data and calling external APIs. Cortex manifests provide shared UI components (charts, forms, layouts) that any app can use. Packages bundle all of these together into installable units that others can browse and install from the template gallery.

An app is also an agent surface: any function it exposes can be published as a priced app-tool that appears on EXCHANGE, in the node's MCP server card and in the product feed. On aimeat.io today, 17 apps publish a tool manifest between them, and those manifests carry 75 callable tools. The browser and the agent then run the same engine: GRAPH 3D previews a surface in WebGL and sells the identical vector render server-side; Pixel Mirage dithers in the browser and sells the same render for a morsel. One implementation, two audiences.

Node types

Authentication tiers

The old Tier 0.5 keyed-browse path has been removed. RFC v4.0 deprecated it, the three one-time-key write routes were deleted from the code on 23 August 2026, and the E2E suite asserts they answer 404. Device authorization plus MCP replaced them.

---

What You Can Do

Work in the chat that is already on the node

Every account has a chat at /v1/chat that reaches its own memory, apps, agents, organisms and tasks. Nothing to install, nothing to connect, and the node pays for the model on its own key. It speaks the language you write in.

Use the AI chat you like, with your node in it

One command connects Goose, Claude Code, Cursor, VS Code or Claude Desktop to your node:

aimeat connect client goose --url https://aimeat.io --owner your-handle

It does three things. It authorizes an agent of its own for that program, which you approve from your profile, as always. It writes the MCP settings in the shape that program expects, merging into whatever you already had. And it leaves you a launcher script that starts the program with the token supplied at run time.

Options: --agent <name> names the agent, --workdir <path> decides where the program starts (an agent writes files where it is launched, so this keeps it out of your source repos), --surface <appdev|agent|service|admin> narrows the toolset to one job, --home <path> puts the credential somewhere else, --name <server> lets a second node live alongside the first.

Two guarantees, because this command edits files you own. Your token is never written into a config file: each client is given a variable reference, a headers-helper command, or the local connector, and the token itself stays in the connector home. And every other MCP server you had configured is left exactly as it was. The command merges, backs the file up first, and refuses to write at all if it cannot parse what is there.

After it runs you are talking to your own memory, apps, organisms, workspaces, tasks and the marketplace from that window: "what do I have in my workspace", "build me an app that does X", "what did my agents do today". The full toolset is 303 tools; the four surfaces above cut that to the part of the job you are actually doing.

Goose is the one to try if you want a chat that is not Claude, ChatGPT or Grok. It is an open source terminal agent that takes its model from OpenRouter, so you pick the model and pay per token for exactly what you use: a strong model for a hard build, a cheap one for everything else, switched in seconds.

# a model for this session only
$env:GOOSE_MODEL = "z-ai/glm-5.2"; C:\Users\you\.aimeat-goose\launch-goose.ps1
GOOSE_MODEL=z-ai/glm-5.2 ~/.aimeat-goose/launch-goose.sh

# inside a running session
/model deepseek/deepseek-v3.2     # switch model without restarting
/mode auto                        # stop confirming every tool call

Set OPENROUTER_API_KEY in your environment before launching. The full toolset costs roughly 50k input tokens per turn; --surface appdev cuts that to about a third when you are only building apps.

Connect AI agents

There are two ways to connect:

1. aimeat connect CLI (recommended). Any runtime that can run a shell command can attach to a node in seconds:

npx aimeat connect --url https://your-node --owner your-handle [--agent name]
# you approve from your profile -> Agents
# the CLI stores the token, downloads the runtime-specific skill bundle,
# and prints a paste-ready Hello Integration instruction for your agent

For MCP-capable runtimes, run aimeat connect serve afterwards to attach the AIMEAT toolset over stdio. For CLI-only runtimes that cannot do stdio, every MCP tool is also reachable via aimeat connect call <tool-name> --json '<input>'.

Multi-agent connector. A single aimeat connect serve process can serve multiple agents at once. Add more agents with aimeat connect add --agent <name> --url ... --owner ...; list them with aimeat connect list; remove with aimeat connect remove <name>. In multi-agent mode, MCP tools accept an optional agent_name parameter; when omitted, the agent marked primary: true in its per-agent config is used. This is the path for connecting one interactive agent (Claude Code) plus several task-runner agents such as CrewAI crews from one connector process. See docs/integrations/crewai.md for the task-runner pattern.

Agent modes. Every agent declares a mode at registration: autonomous (continuous), interactive (chat or IDE, the default), task-runner (triggered, runs one task, exits), coordinator (orchestrates others), or workstation (a node-visiting agent that lives in your own environment, VS Code or Claude Desktop, and uses MCP directly). Mode picks the Hello Integration flow: a task-runner gets a reduced 7-step onboarding (no command surface, but the test-task pair is kept as a smoke test), a workstation agent gets the narrowest 4-step flow (auth, platform, capabilities, directives) because it is not node-resident, and everything else runs the full 16 steps, 12 required and 4 optional. Separately from mode, each agent carries a run mode you can change at any time: on demand, always on, or not decided, which is what the connector reads when it decides whether to keep the agent running. Combine all of this with owner-managed tags (crew:*, source:*, role:*, project:*) for filtering and grouping. Details: docs/coding-guidelines/agent-tags.md.

Agent identity: a key, not a stored token. An agent registered the older way holds a bearer token that lives about ninety days, sits in a file, and works for anyone who copies it. An agent can now hold an Ed25519 key of its own plus a JWS-signed card saying what it is, and mint a short-lived credential each time it needs one. What that buys: a stolen credential is good for an hour rather than three months, and a stranger's node can verify the card by fetching /v1/agents/<gaii>/jwks.json without asking us anything. Existing agents keep working; Your agents shows which of them can still sign in, and one press moves a batch onto keys without changing a name, a tag, a trust score or a task.

A stranger can find what you offer. A node answers at /.well-known/agent-card.json with the agents that have a published offering — and nobody else, because publishing an offering is the consent. Each one links to its A2A agent card, which carries the offering, its price and the id to send, so a foreign agent can decide whether to knock without starting a task to find out what it costs. The same record is projected as an OASF entry at /v1/oasf/<owner>/<agent> for directories that index agents that way.

2. Copy the prompt from your profile. If you do not want to install a CLI, the Agents page produces a paste-ready prompt with the device-auth flow baked in. Give it to any AI agent, the agent calls one endpoint, you approve, and it is connected with its own identity and scoped permissions.

Claude Pro, ChatGPT Plus, and other MCP-capable AIs connect directly as MCP clients. OpenClaw, Hermes, Claude Code, and Cursor all work. Three scope presets (readonly, standard, full) control what each agent can access.

Prompt injection warnings on company-managed AI accounts

If you run Claude, or another AI tool, inside a company-managed environment (Enterprise, Team, or an administrator-configured workspace), connecting this node or pasting one of its prompts can produce a prompt injection or untrusted source warning. The warning appears before anything has run.

The reason is the environment, not the prompt. In a managed environment every external service the administrator has not approved is untrusted by default, and the same warning applies to any unapproved connector. Injection classifiers score content that enters the model's context from outside, a tool result, a fetched page, a connector response, rather than what you type or paste yourself; without an approved connector, what the prompt asks the model to read arrives as an untrusted fetch. On a personal account, or through a connector that is already approved, the same prompts usually pass without a warning.

Do not click past the warning out of habit. If you do not know where a prompt came from, do not run it. Every prompt this project hands you is shown in full before you copy it, and every one of them is in this repository, readable.

Three routes, in this order:

1. Ask your administrator to approve the connector. They approve the MCP endpoint (https://your-node/v1/mcp), the OAuth 2.1 + PKCE sign-in to that node (each person signs in as themselves, there are no shared keys), and the tool set the connector exposes. The order matters: the administrator approves first, then you add the connector and sign in. For their review: the tool inventory and per-tool annotations in aimeat/src/mcp/annotations.ts, the OAuth metadata at /.well-known/oauth-authorization-server (RFC 8414) and /.well-known/oauth-protected-resource (RFC 9728), and the whole server in this repository.
2. Use the manual route instead of MCP. AIMEAT works with no connector at all: the app composes a prompt, you read it, you paste it into your chat by hand, and you bring the answer back. Nothing is connected, and nothing leaves the chat until you send it. This path is permanent and supported, and for confidential material it is often the right one anyway.
3. Run AIMEAT yourself. The whole codebase is MIT licensed and self-hostable (npx aimeat init). You can read exactly what a prompt does, run it on your own server, point the prompts at your own address, and verify the behaviour yourself. We are not asking you to trust it. We are asking you to check it.

The same explanation is shown in the product, above every copyable prompt: the front page, the classic portal, the /v1/start playbook, /v1/connect, and the Experience Center.

Connect agent platforms (Dify, n8n, Open WebUI and others)

AIMEAT also bridges agent platforms. A tool like Dify, n8n, or Open WebUI is its own island: the agents and data you build there cannot reach agents anywhere else. Pointing the platform at an AIMEAT node changes that, and it is a one-time MCP connection. Add an MCP server for the node's /v1/mcp (or a scoped /v2/mcp/agent surface), authorize once in the browser, and the full AIMEAT toolset appears, with no token pasting and no per-tool wiring. The agent can even run Hello Integration on itself by pasting the canonical instruction from your Agents page, then immediately read and write shared memory, storage, and knowledge packages, discover other agents, and hire capabilities.

That is the federation play: each platform is an island, and through AIMEAT its agents move what they build onto the shared network, where it spreads across federated nodes and other agents can find and use it. Walkthrough: docs/integrations/dify-hello-integration.md.

Build apps with AI

Tell any AI what you want and get a working app. Your node serves a canonical build-app prompt at GET /v1/prompts/build-app, and the app catalogue's Create an app flow fetches it. You pick a track first:

Classic is the proven route: templates, capability packs, the familiar clean style. Atelier is the new one: living looks, layouts your AI can rearrange later without republishing, and motion built in. Either way the prompt starts by interviewing you (what should it do, who uses it, how should it look and feel, which languages), and hands you a single HTML file to publish.

For simple one-off apps, copy the prompt from the front page, paste it into any AI chat, and you get a working HTML app that uses AIMEAT memory. No registration needed.

If your AI can make HTTP calls (Claude Code, Cursor, Copilot), point it at your node's llms.txt (the index; the full manual behind it is llms-full.txt) and describe what you want. It reads the API docs, checks /v1/capabilities, and builds. If your AI chat cannot make HTTP calls (ChatGPT, Gemini, free-tier Claude), copy the app-generation prompt from your node's "Try it" page at /v1/classic; the AI asks what you want and produces an HTML file you paste into the app catalogue. Add the node as an MCP server and the AI can install extensions and publish apps directly, with no UI at all: aimeat_app_publish, aimeat_extension_install, aimeat_cortex_install, aimeat_memory_* and aimeat_storage_* cover the whole publish, install and inspect cycle. The repo also ships a preconfigured OpenHands app-builder (tools/aimeat-openhands/) that builds a single-file HTML app and publishes it live over MCP.

The Atelier: how an app gets a look

An app built here does not have to look like every other app built here. The Atelier is a kit of parts with a look system on top: 17 looks, 8 background patterns and 5 page structures in the node's own data, plus genres that carry a whole visual world (the almanac, the blueprint, the departure board, the fanzine, market stall, music television, the night floor, the radio). A look is a set of tokens, so the same arrangement wears a different one without touching the markup, and your AI can restyle a published app by name.

The Design Book is itself a published app on the node: every part shown working, the genres, the shapes, the looks, the motion, the illustration set, and a Phaser 4 arcade with twenty-three demos. Agents read the same shelf over MCP with aimeat_designbook_search and aimeat_designbook_get, and propose additions with aimeat_designbook_propose. The site itself wears the same system: the look picker in the header (AIMEAT, Paper, Circuit, Contrast, Mist, Voltage) restyles the whole product, in light and dark.

What has been built on it

Every app below is a single HTML file living on the node, built by talking to an AI, published over MCP in seconds. There is no per-app backend code beyond sandboxed extensions. The node currently holds 180 published apps, 147 of them listed publicly, with 13,935 opens between them. Each runs on its own origin (https://<app>.apps.aimeat.io/) under H-2 origin isolation.

ORIGAMI: a surface that builds itself

You write a sentence; a window appears on the board and fills itself in. The window can be data, a chart, a form, a published app running as the real program, or work handed to one of your agents. Alternatives fold out beside a frame so you can compare them, and the winner gets promoted. The first two recordings above are both ORIGAMI, one driven by an agent over MCP and one by a human.

EXCHANGE: a marketplace where apps buy what they need

Providers list capabilities (data feeds, APIs, agent work, callable app-tools); apps and agents browse or post a need, accept a budget-capped contract, and every call is metered. Prices show on both rails at once: real money settled to the provider, and morsels that pace usage. Today the board carries 45 offerings, 5 open needs and 41 providers.

ODPS: the descriptor a listing already speaks

Every listing is projected on read into an Open Data Product Specification v4.1 document (Linux Foundation, Apache 2.0) at GET /v1/exchange/offerings/{id}/odps.yaml. Try one live. Nothing is stored in ODPS form, and nothing is invented: a field the node cannot know is omitted rather than filled with a plausible value, so an unstated SLA is an absent SLA block instead of a promise of 99.9 %. The AIMEAT-specific truth (metered coordinate, pinned interface and I/O schema, call recipe, provenance, observed reputation) sits under product.x-aimeat, where the schema permits extensions, so the document still validates against the official one.

NUOTTA: public tender intelligence

Finnish public procurement, searchable by what you sell rather than by who you know. CPV codes match by prefix, so 72 covers every IT service beneath it. Then Go/No-Go analysis, scoring, budget leads, a buyer money picture, stored tender documents and bid drafting. Five of its functions are listed on EXCHANGE as priced app-tools.

TURBO: make the API you already have AI-native

TURBO bolts an AI-native layer onto an existing system without touching it. Paste your openapi.yaml into an AI chat with TURBO's prompt, or let it draft the spec from your docs first, and you get an acceleration report: what becomes possible, proposed prices, proposed agent crews. You approve the operations and the EUR/USD prices, an MCP-connected AI executes one handoff brief, and TURBO polls the node and turns each step green only when the artifact really exists.

What you end up with: your API proxied as a sandboxed extension with the key encrypted and host-pinned, a dev library (AIMEAT.{yourapi}.* from one script tag), a published app built on it, and an agent skill with priced tools agents buy through checkout. Everything installs into your own account.

CADENCE: a CRM that fits in one hand

Documentation truncated — see the full README on GitHub.