Programmable vision for AI agents. Request live video or single frames from real-world locations, settled on-chain in USDC.
This repo contains:
Package | Purpose |
TypeScript HTTP client for the Luxxon API. Pure | |
MCP server exposing Luxxon as agent-callable tools (stdio transport, works with Claude Desktop / Cursor / any MCP client). | |
Runnable agent examples — copy, paste, run. |
API: docs.luxxon.dev — https://api.luxxon.dev/api/v1.
Agents have fetch(), not <video> tags. The MCP server turns Luxxon
into a set of tools any MCP-aware agent can call directly:
get_frame(sessionId) → JPEG bytes ready for a vision model
get_session(sessionId) → current state of a live session
cancel_session(sessionId) → early termination
Drop the server into your Claude Desktop config and your agent can see the world.
Related MCP server: StatePulse API
import { Luxxon } from "@luxxon/sdk";
const lx = new Luxxon({ apiKey: process.env.LUXXON_API_KEY });
// Read the latest frame from a LIVE session (returns a Buffer).
const jpeg = await lx.sessions.frame("sess_abc123");// ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"luxxon": {
"command": "npx",
"args": ["-y", "--package=@luxxon/mcp", "luxxon-mcp"],
"env": { "LUXXON_API_KEY": "lxxn_live_..." }
}
}
}Restart Claude Desktop; the three tools appear under "Luxxon".
v0, building in public. The SDK + MCP cover the agent-read surface today (frames, session state, cancel). Session creation lands behind a custodial-wallet sign path that's still being built on the API side — for now you create sessions yourself via the REST API and use this SDK/MCP to consume them.
Public dev console + Python/Go/CLI SDKs follow. Watch this repo.
Cancel a pre-LIVE Luxxon session (REQUESTED or ASSIGNED). For LIVE sessions, end the session via the REST API or a future end_session tool — this one will error with INVALID_STATE.
Inputs: sessionId (required).
Fetch the latest decoded video frame from a LIVE Luxxon session as a JPEG image. Hand the returned image straight to a vision model. Returns 404 FRAME_NOT_AVAILABLE for ~3-5s after /start while the first keyframe arrives — retry if you hit that.
Inputs: sessionId (required).
Read the current state of a Luxxon session (REQUESTED, ASSIGNED, LIVE, ENDED, etc.) plus meter fields (cleanSeconds, chargedMicroUsdc, settlementTxHash).
Inputs: sessionId (required).
Return a WHEP playback URL for a LIVE session. For agents with their own WebRTC stack; most agents should prefer get_frame instead.
Inputs: sessionId (required).
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 4 tools
Each tool has a distinct purpose: cancel_session cancels pre-LIVE sessions, get_frame fetches a video frame, get_session reads session state, and get_stream_url provides a playback URL. No overlap exists.
All tools follow the verb_noun pattern (cancel_session, get_frame, get_session, get_stream_url) with consistent snake_case, making the naming predictable.
At 4 tools, the set is slightly small but still reasonable for a focused domain. It covers basic read operations and one mutation, though some common operations are missing.
The tool set is notably incomplete: it lacks end_session for LIVE sessions, forcing agents to use a REST API. The stated dead-end for cancel_session on LIVE sessions creates a significant gap in the domain lifecycle.
Pay-per-use MCP tools for AI agents. USDC on Base via x402. No API keys.
17 pay-per-call web, finance, legal and onchain tools for AI agents. USDC on Base via x402/MPP.
63 pay-per-call tools for agents: vision, text, data, web, blockchain. USDC on Base via x402.
Pay-per-call tools for AI agents: product photos, documents, web pages, data. USDC via x402.