MCP setup and agent work

flowchart LR
  A[Agent inspects project] --> B[Plan selected stages]
  B --> C[App language-model provider]
  B --> D[Host agent text work]
  C --> E[Reviewed artifacts]
  D --> E
  B --> F[ASR, TTS or render as needed]
  F --> E
  E --> G[Inspect and export results]

Managed HTTP setup

Install or enable the MCP component in Manager, start it, then copy the generated configuration for your client. The usual local endpoint is http://127.0.0.1:8099/mcp. Use the dedicated private bearer credential supplied for that connection. Do not publish it or paste it into a shared conversation.

Restart or reconnect the client and ask it to inspect system status. It should see the intended instance and existing projects before creating new work.

Stdio alternative

For an existing configured target, pandrator-mcp doctor --target my-target checks the connection. pandrator-mcp host-config codex --target my-target prints a stdio host configuration without embedding its secret. Replace my-target with the actual configured target name.

How work runs

In a native workflow, the configured Pandrator language model performs applicable text stages. In a passive workflow, the host agent can read bounded text and save reviewed proposals or results. Recognition, speech generation and rendering still run through their engines and jobs.

The agent should inspect existing work, plan the next operation, review required confirmations, execute once and monitor the returned job. MCP does not make every long task instantaneous or remove the need to review generated text and audio.

Transcript-only requests

pandrator_transcribe can return a temporary result without creating a project. Save it when ready: results expire one hour after completion. The limit of twenty retained temporary transcriptions per owner includes finished results. For a batch, ask the agent to save the transcripts, then release only its own completed temporary results with pandrator_transcription_delete. Use a project when you need correction, translation or later revisions.

Check the connection with a small task

Ask the assistant to identify the connected Pandrator instance and list the existing projects before creating one. Then try a short transcript or one narration passage. This checks the workflow you intend to use, beyond the server's health response.

Keep the work handle and the selected artifact or generation-run ID in the conversation's task record. After a reconnect, inspect those same IDs to continue. A passive text run still needs the host agent to claim, review and submit its text batches; creating the run alone does not process them.

See the MCP tutorial.