Install the local integration
Follow getting started to install the CLI and sign up. Then run inside your repository:
cscale init
Init installs the pre-push hook. It adds MCP, the skill and supported harness hooks only where .claude or .codex exists at the repository root. To choose an integration explicitly, run cscale init --for claude-code or cscale init --for codex. To preview the files for Codex:
cscale init --for codex --dry-run
The init reference lists the supported harness choices. Restart or reconnect your agent after installing its integration.
Connect another MCP client
Configure the client to launch this process over standard input and output:
cscale mcp
The process stays attached to the MCP client. Send one JSON-RPC message per line on standard input. This example initializes the session, lists the tools, and calls whoami:
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"example","version":"1.0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}
{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"whoami","arguments":{}}}
Wait for the initialize response before sending the notification, and for the list before calling a tool. Responses are one JSON message per line on standard output; match them by id. The notification has no response. The live tools/list is authoritative for the account. The MCP reference describes the connection options available on your plan and gives tool names, input schemas and examples.
From CLI 0.2.10, successful whoami and get_namespace_plan results are in result.structuredContent, like the other tools. CLI 0.2.9 returned them as JSON in result.content[0].text, so parse that text if you still run 0.2.9. The tests_check report is in result.structuredContent as described below.
get_namespace_plan needs the person's sign-in. Ask them to run cscale auth login --email you@example.com, open the email link and type the code shown in the terminal. A check-only key cannot read the plan. The API reference lists each operation's credential requirements.
Run the agent loop
Ask the agent to check its work after committing. It should call tests_check, read the findings, fix the tests or code, commit, and run a full check again. For example, the tool call uses:
{
"name": "tests_check",
"arguments": {"base": "origin/main", "head": "HEAD"}
}
Focused results do not authorize a push. The tests_accept tool records an agent's reason for an acceptable finding; the target's policy decides whether it counts. A required passkey confirmation stays with the person. When the check stops for sensitive files, the agent asks the person before it continues.
The result is a version 1 report object in MCP's structured content. It has a schema version, check ID, scope (full or finding), state, checks, findings, problems and a gate. The gate's gate.status, gate.exit_code and gate.summary say the policy outcome; each finding has an id, kind, blocking, summary and remedy, along with its path and evidence. Read the findings even when the gate passes: the target's policy decides what blocks. A focused result is not a full check.
The first CScale 1 question can take up to a minute while it starts; later questions are quicker. Allow at least a minute for each model request, and more for the complete check.
Start an account from an agent
An MCP client can start sign-up for a person. Call signup_start, ask the person to confirm the email link, then call signup_wait with the returned local attempt ID. The CLI keeps the issued check key private. It is not part of the tool result.
Use tests_usage to read model usage and contact_support to send a support message. The API reference covers integrations that call the server directly.