Connect Claude Code to AdaptOrch over HTTP MCP.

Create an account at /app/signup and an API key at /app/api-keys. Model execution also needs your provider, model and provider key; provider usage is billed by the provider separately from the AdaptOrch plan. Set these only in your local environment, never in a shared transcript or committed file.

export ADAPTORCH_API_KEY="ado_..."
export ADAPTORCH_PROVIDER="openai"
export ADAPTORCH_PROVIDER_MODEL="YOUR_SUPPORTED_MODEL"
export ADAPTORCH_PROVIDER_KEY="YOUR_PROVIDER_KEY"

Run the command in the shell where Claude Code is installed. Claude Code stores the resulting headers in its MCP configuration, so treat that configuration as private.

claude mcp add --transport http adaptorch https://adaptorch.com/mcp \
  --header "Authorization: Bearer ${ADAPTORCH_API_KEY}" \
  --header "X-Provider: ${ADAPTORCH_PROVIDER}" \
  --header "X-Provider-Model: ${ADAPTORCH_PROVIDER_MODEL}" \
  --header "X-Provider-Key: ${ADAPTORCH_PROVIDER_KEY}"

The shell expands the four variables when the command runs. The provider headers are required for model execution; the AdaptOrch key alone authenticates the service but does not supply model credentials.

This read-only tools/list request checks the endpoint and your AdaptOrch key. It does not submit a run or spend provider budget. A valid key returns a result.tools array that includes adaptorch_run. Without a key the endpoint answers HTTP 401 with Missing bearer token, which still proves it is reachable.

curl --fail --silent --show-error https://adaptorch.com/mcp \
  -H "Authorization: Bearer $ADAPTORCH_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Ask Claude Code to call adaptorch_run once with the single-task payload below. This consumes plan allowance and provider budget, so do not resubmit it while polling.

{
  "payload": {"subtasks": [{"id": "ping", "description": "Reply exactly PONG", "estimated_tokens": 16}]},
  "synthesis_mode": "paper",
  "budget_policy": {"global_timeout_budget_seconds": 20}
}

Use the returned run_id with adaptorch_get_run and adaptorch_get_artifacts, or open /app/runs.

The run record shows the inputs, the selected model and outputs, recorded commands and check outcomes when an execution environment is configured, and open review items. A SUCCEEDED run is not a test pass: a repository code check needs a separately configured execution environment, and missing or skipped checks stay unverified.

AdaptOrch documents the Claude Code configuration format above, but this connection has not been tested by AdaptOrch end to end in Claude Code. Read the claim boundary before relying on run records in a review process, and see the AI PR review guide for what to inspect.

See all AdaptOrch integrations, or the full MCP documentation for Gemini CLI, VS Code and Windsurf configuration formats.