Skip to main content
The standard @crowdlisten/harness@2.4.3 installation combines the native record tools with authenticated hosted expert tools. The native tools remain:

Expert tools in the same connection

The signed-in standard client discovers the currently deployed schemas. Key tools include: Read the tool schemas for approval, status, context selection and video operations; the set can grow without another package installation. Setup never grants permission for paid work. Use the existing task ID after an uncertain response instead of creating a replacement job. Set up your client before making requests. The installed tool schemas are the complete contract for supported modes and fields.

Hiring MCP for ChatGPT and Grok

The hiring profile uses https://crowdlisten-skills-api.vercel.app/skillshare/api/mcp/hiring, Streamable HTTP, OAuth with PKCE, and the experts:hire scope. It supports a focused expert-hiring toolset rather than all local harness tools. The public OAuth endpoint is enabled for testing; account connection and end-to-end host acceptance remain under verification. See ChatGPT setup or Grok setup for availability, authorization, quote approval, results, and disconnect instructions. These use a separate endpoint from the legacy native-tools service below.

Legacy remote MCP

The remote endpoint is https://mcp.crowdlisten.com/mcp. This legacy host exposes only the three native tools; it does not yet have parity with the unified standard package for hosted experts, video and context. The client must securely supply a valid CrowdListen account session using an Authorization: Bearer header and manage its refresh. The endpoint currently does not implement OAuth or anonymous URL-only setup. A successful initialize or tools/list response does not establish authorized data access. Verify recall({ mode: "connection" }) from the connected client. Never paste session credentials into chat. Most local clients should use the browser-authenticated stdio harness in the setup guide.

Expertise

The expertise studio and its creation/evaluation flow remain in development previews. Enabled pilot accounts can create and review a private method there. Choose an authorized workspace, then read the method and its revision:
Quote the input before running it. Quotes do not run a model or debit credits.
Check ready, maximumCredits, availableCredits, expiresAt, and quoteHash. After the user authorizes that scope and maximum charge, submit the same revision and files:
Preserve the request ID if the response is uncertain. Read progress rather than starting another task:
The managed task continues after the client disconnects. Retrieve the published artifactId and exact file revision from its manifest:
Result reads are free. Use the returned revision and a listed path; do not guess filenames. These reads do not launch a new analysis or change sharing. Cancel with:
Work already performed can consume credits. Uncertain usage remains reserved until reconciled. Workspace access and method ownership are checked on each call.

Research and collection

Inspect subjects with recall({ mode: "entities" }). Verify a chosen subject using recall({ mode: "connection", entity_id: "ENTITY_UUID" }). This checks current scoped reads and permissions; it does not prove live platform access. recall({ mode: "collection_plan", entity_id: "ENTITY_UUID" }) reads the saved source scope. An agent with authorized browser or source tools can collect originals, save them with ingest({ destination: "sources", ... }), and record coverage separately. See the collection skill for the complete schema and retry rules.

Reviewing without expert tools

The standard local package supports the full expert toolset. Hosted assistants can instead use the browser-approved account connection to read saved results. That read-only HTTPS connection does not start jobs, spend credits or submit result decisions. It is separate from the legacy remote MCP endpoint. An owner can provide a selected-file share or review export for read-only feedback. This grants access only to the shared material, not other workspace context or expert_result_review. The owner applies any requested changes through their authorized agent. Discovery, installation, and reading a supplied review export must not start paid work.