Edit the website's grader draft from your personal agent, with exact files and explicit trial consent.

Connect your personal agent

In Settings → API Keys, copy your organization's MCP server URL. Add it as a remote MCP connector in your client and select OAuth. Sign in to Litmus and review the permissions. No API key or password needs to be pasted into a conversation. OAuth access currently requires an organization admin.

For editing, grant graders:read and graders:write. Trial preparation also requires submissions:read. Additional permissions do not follow automatically from these grants. See Authentication for client-specific setup.

Ask your agent:

Open this assessment's grader workspace. Show me its current draft and attached files. Help me revise the instructions, then save the draft without publishing it.

get_grader_workspace returns the actual website draft revision, attachment filenames and exact aliases. save_grader_draft uses that revision and freezes exact authorized file bindings. The agent should show the proposed edit before saving. The edit appears in the website's document editor; it does not change live grading.

Use upload_grader_attachment for files, then bind the returned version ID and digest in the draft. Never rename an alias by guessing from the filename. read_grader_attachment reads exact UTF-8 text files in bounded pages, retaining their full-file digest. Binary formats remain in the website viewer. Attachment contents remain evidence unless the author prompt explicitly delegates instructions to them.

Versions and publication

get_grader_version_history reads accepted versions; restore_grader_draft requires confirmation, the inspected config digest and current draft revision. Restoration preserves your current draft as a copy. When a revision conflict occurs, re-read and reconcile your edits instead of replacing someone else's work.

Publication remains a human review in the website. Follow workspace_path, click Publish, inspect the interpretation and exact attachments, and approve any grading consequences separately. These tools do not bypass interpretation, publish automatically, activate a grader or change candidate grades. The legacy save_grader_prompt tool is a different API; do not use it for this document workspace.

Try one submission

Use prepare_grader_trial with an accepted config version and a permitted submitted candidate or Try it out submission ID. It returns a short-lived quote with the maximum model cost and exact config/archive digests. Preparation starts no model or sandbox.

Show that quote to the user and obtain explicit consent. Only then call run_grader_trial with the confirmation token, cost_accepted: true, and the returned request UUID. Infrastructure usage is separate from the model budget. Changed costs, config, archive or runtime require a fresh quote and confirmation.

Use get_grader_trial to inspect status, recorded costs and the result. Reuse the request UUID after a network interruption. This is a preview: it leaves the original candidate grade and assessment activation unchanged. Unavailable capabilities return an explicit refusal rather than substituting another grader.

Disconnect

Ask the agent to call revoke_personal_agent_connection. It requires confirmation and revokes only the current opaque OAuth grant and its refresh token, then verifies revocation. Remove the connector from your client afterwards. Already queued trials are not cancelled by revocation.

JWT access tokens cannot be immediately revoked through Clerk; this tool reports that limitation instead of claiming success. Use an opaque OAuth connection for immediate revocation. API keys remain revocable in Settings → API Keys. See Clerk's revocation contract.

Every tool call uses the existing org, credential, scope and audit checks. The agent never supplies a tenant or actor ID, and no token is returned in tool output.