THRONE
run a server

methodology

Calibrated checks, raw traces, blunt verdicts.

Throne does not host your server or build a client. It exists to answer one release question: does this MCP server survive the live clients developers actually use?

the run sequence

Every run follows the same route.

  1. 01 / connect to the repo, package, command, or config file
  2. 02 / discover tools, resources, prompts, and declared capabilities
  3. 03 / validate tool names, schemas, defaults, and required fields
  4. 04 / smoke-test every tool with known-good inputs
  5. 05 / force error handling paths and inspect surfaced failures
  6. 06 / test streaming behavior and partial result handling
  7. 07 / test resource lifecycle, subscriptions, and cleanup
  8. 08 / run concurrent calls and timing-sensitive operations
  9. 09 / reconnect and verify recovery behavior
01

DISCOVER

Start from the artifact

Throne reads the config, package metadata, or GitHub repository and starts the server as a user would.

throne run github.com/acme/files-mcp
02

EMULATE

Apply client behavior

Calibrated emulators model Claude Code, Cursor, and ChatGPT Desktop behavior around the parts that break releases. Every emulator is verified weekly against the real client. An emulator that has not passed parity is labeled calibrating and produces no verdicts.

live_clients = ["claude-code", "cursor", "chatgpt-desktop"]
calibrating = ["codex-cli", "zed"]
03

TRACE

Record what happened

Every result points back to raw JSON-RPC requests, responses, timings, and the emulator rule that produced the outcome.

{ "jsonrpc": "2.0", "method": "tools/call", "elapsed_ms": 812 }
04

SCAN

Inspect the security surface

Static rules inspect dangerous path, shell, prompt, secret, rate, and dependency patterns.

THR-PATH-01 HIGH / escaped declared root
05

DECIDE

Make the release call

The report ends with a verdict: fit to ship, needs review, or blocked before release.

verdict: NOT FIT TO SHIP

vendor neutral

Throne is not affiliated with Anthropic, Cursor, OpenAI, or Zed.

The point is independence: one release gate that measures client behavior without selling you a client, IDE, hosting layer, or framework.