OculusMind.AI

Fournit un accès MCP aux agents OculusMind.AI qui recherchent sur le web, analysent les actions et les produits, automatisent et planifient des workflows et utilisent d'autres outils de serveur MCP.

GitHub

Documentation

OculusMind Agents

Open-source tooling to talk to OculusMind.AI agents — Uno, Vera, and Maximus — from your own code or from any MCP host.

Two ways in:

MCP server — drop the agents into Claude Code (or any MCP-speaking host) as tools. Ask Uno/Vera/Maximus a question straight from your CLI. See servers/mcp.
A2A client libraries — call the agents directly from your program over the open A2A protocol (send a message, stream a reply, discover capabilities). Available for TypeScript and Python.

Both speak the same wire contract (docs/PROTOCOL.md) and use the same authentication — a bearer token you mint from your account (docs/AUTHENTICATION.md).

MCP vs. A2A. These are different standards, and this project bridges them. The agents themselves are served over A2A (Agent-to-Agent). The MCP server here wraps those A2A endpoints as MCP tools so an MCP host can call them; the A2A client talks to them natively. Pick whichever fits.

Quickstart — MCP server in Claude Code

Mint OCULUSMIND_TOKEN at https://oculusmind.ai/integrations, then:

claude mcp add oculusmind \
  --env OCULUSMIND_TOKEN=<your-token> \
  -- npx -y @oculusmind-ai/mcp-server

Or run it from a clone instead

# npm install also builds the in-repo A2A client this depends on
cd servers/mcp && npm install && npm run build

claude mcp add oculusmind \
  --env OCULUSMIND_TOKEN=<your-token> \
  -- node "$(pwd)/dist/index.js"

Then in Claude Code: "Use oculusmind to ask Vera to verify the specs of the DJI Mini 5." The ask_agent and list_agents tools appear as mcp__oculusmind__*.

Quickstart — A2A client (TypeScript)

import { OculusMindAgents, StaticTokenProvider } from '@oculusmind-ai/agents';

const client = new OculusMindAgents({
  agent: 'vera', // or 'uno' | 'maximus'
  tokenProvider: new StaticTokenProvider(process.env.OCULUSMIND_TOKEN!),
});
const reply = await client.send('What is the weather in Austin?');
console.log(reply.text);

Quickstart — A2A client (Python)

from oculusmind_ai_agents import OculusMindAgents, StaticTokenProvider

client = OculusMindAgents(
    agent="vera",  # or "uno" | "maximus"
    token_provider=StaticTokenProvider(os.environ["OCULUSMIND_TOKEN"]),
)
print(client.send("What is the weather in Austin?").text)

Getting a token

Sign in at https://oculusmind.ai.
Open the Integrations page.
Mint a bearer token (valid 7 days) and copy it.
Provide it as OCULUSMIND_TOKEN (or your own secret store).

Full details, security guidance, and the token-refresh plan: docs/AUTHENTICATION.md.

Layout

├── servers/
│   └── mcp/               MCP server for Claude Code & other MCP hosts
├── clients/
│   ├── typescript/        A2A client library (@oculusmind-ai/agents)
│   └── python/            A2A client library (oculusmind-ai-agents)
│                          The MCP server consumes the TypeScript client
│                          directly from this repo — one A2A implementation,
│                          not two. Wired as an npm workspace; nothing is
│                          fetched from a private registry.
├── docs/
│   ├── AUTHENTICATION.md  how to get a token + the refresh best-practice
│   ├── PROTOCOL.md        the A2A wire contract both speak
│   └── ROADMAP.md         languages, token auto-refresh
├── examples/              runnable samples
└── LICENSE · NOTICE

Development

This repo is a single npm workspace. From the root:

npm install
npm run check              # build + lint + typecheck + format + unit tests
npm run test:integration   # spawns the built MCP server over stdio

# Live tests against production are opt-in; without a token they self-skip.
OCULUSMIND_TOKEN=<your-token> npm run test:integration

Never commit a token — the integration tests read OCULUSMIND_TOKEN from the environment, and .env* files are git-ignored.

Status

Early. The TypeScript and Python clients and the MCP server are the reference; the protocol and auth contracts are stable enough to build against. Breaking changes are possible until 1.0.