MCP Server

Ad Astra runs a public Model Context Protocol server at mcp.adastracomputing.com. Read tools take no auth; one write tool submits a contact intent on a user's behalf. It also exposes search and fetch, so it works as a ChatGPT Deep Research source. Run any read tool below and inspect its live structuredContent.

Live verifier

Try the tools

Each card below fires a real tools/call against mcp.adastracomputing.com from this page. No server-side proxy.

Endpoint

• URL: https://mcp.adastracomputing.com
• Transport: Streamable HTTP, JSON-RPC 2.0
• Protocol version: 2025-06-18 (also accepts 2025-11-25 and a 2025-03-26 backwards-compat fallback)
• Discovery: apex · server

Write tool

submit_intent sends a structured contact intent (demo_request, consulting_inquiry, partnership, press, research_collab) to the lab. Requires contactConsent: true asserting the user authorized the message. Returns a signed receipt the agent can present to its principal. Not exposed in the playground; it is a real write that emails the lab inbox. Use it from a real MCP client.

Auth and limits

No auth on read tools; CORS allows browser callers. Per-IP and per-Origin rate limits apply to submit_intent. Read tools share a soft per-IP ceiling. Repeated abusive traffic fails closed.

Receipts

The signed receipt from submit_intent binds the canonical payload hash, the issued-at timestamp and the intent type. It does not carry the auditId, and a valid receipt therefore proves the worker minted it, not that the audit row was written. Receipts issued from 19 August 2026 are Ed25519 over the key published at did:web:mcp.adastracomputing.com, so anyone can verify one independently, the same guarantee as INK envelopes. Older receipts are a symmetric HMAC that only this server can check. Either way, call the verify_receipt tool, or paste it into the panel on the home page, to see which scheme applies.

Data handling

The audit table stores a one-way SHA-256 hash of source IP, payload and receipt. No raw IP, no raw payload, no raw receipt in the audit row. The agent's contactConsent flag is the lab's assertion that the human authorized the message. See /privacy.

Implementation

Cloudflare Worker, Cloudflare D1 for the audit log. INK envelopes are signed Ed25519. The same worker also serves the INK Agent Card at mcp.adastracomputing.com/.well-known/ink/agent.json for agent-to-agent traffic over the INK protocol. INK spec at ink.tulpa.network/spec.

Troubleshooting

The verifier above tells you whether the apex manifest, the server manifest and the live tools/list agree. If a tool returns a non-200, the meta row on the card shows status and timing. The server treats malformed JSON-RPC as -32700, unknown methods as -32601, and tool errors as a JSON-RPC error with a textual message.

Support

Open an issue on a relevant repo under github.com/Ad-Astra-Computing or email [email protected].