Use Starlight from your own agent code
Connect an agent built with the Vercel AI SDK, the Claude Agent SDK, or Vercel Eve to the Starlight MCP endpoints.
starlightintelligence.ai/app/mcp/, docs-platform/app/api/mcp/route.tsIf you are building your own agent in TypeScript or Python rather than driving a CLI harness, the same Streamable HTTP MCP endpoints work directly from code with zero API keys required.
What was tested: On 21 October 2026 the Vercel AI SDK example below connected to
https://starlightintelligence.ai/mcpusing@ai-sdk/mcp@2.0.79and discovered all eight tools. The/api/mcpendpoint on this docs platform speaks the same JSON-RPC 2.0 HTTP protocol.
Vercel AI SDK (@ai-sdk/mcp)
Install ai and @ai-sdk/mcp. createMCPClient converts the remote MCP tools into typed AI SDK tool definitions you can pass directly to generateText or streamText. Always close the client in a finally block.
import { createMCPClient } from '@ai-sdk/mcp';
import { generateText, isStepCount } from 'ai';
const starlight = await createMCPClient({
transport: { type: 'http', url: 'https://starlightintelligence.ai/mcp' },
});
try {
const tools = await starlight.tools();
const { text } = await generateText({
model: 'anthropic/claude-sonnet-5.5',
tools,
stopWhen: isStepCount(5),
prompt:
'Plan an install of growth-studio for Codex and summarise the verification and rollback steps.',
});
console.log(text);
} finally {
await starlight.close();
}Claude Agent SDK
Register the endpoint under mcpServers and allow its read-only tools via allowedTools. Tool names follow the pattern mcp__<server>__<tool>, so mcp__starlight__* permits all eight Starlight tools without granting broad shell or file permissions.
import { query } from "@anthropic-ai/claude-agent-sdk";
for await (const message of query({
prompt: "Search Starlight Atlas for memory and cite the sources you used.",
options: {
mcpServers: {
starlight: { type: "http", url: "https://starlightintelligence.ai/mcp" },
docs: { type: "http", url: "https://docs.starlightintelligence.ai/api/mcp" },
},
allowedTools: ["mcp__starlight__*", "mcp__docs__*"],
},
})) {
if (message.type === "result" && message.subtype === "success") {
console.log(message.result);
}
}Use allowedTools rather than bypassPermissions so only the read-only MCP tools are auto-approved.
Vercel Eve
In Eve (agent/connections/), each file defines one MCP connection named after the file:
// agent/connections/starlight.ts
import { defineMcpClientConnection } from 'eve/connections';
export default defineMcpClientConnection({
url: 'https://starlightintelligence.ai/mcp',
description: 'Starlight: read-only package plans, Atlas research, and team plans.',
});Guidelines for production callers
- Plans vs. side effects: The tools return deterministic plans and source-backed records. Any action that mutates state (such as installing a package) remains your application's responsibility.
- Rate limits: Cache catalog and documentation responses where practical; public endpoints allow 60 requests per minute per IP.
- Full schemas: See the MCP tools reference for every input field and constraint.