Skip to main content
Get a usable speech file from your coding agent, then transcribe that same file and compare it with the original sentence. This walkthrough follows the local MCP → saved WAV → CLI transcription path completed in both Cursor and Claude Code. It includes the setup and playback interventions observed in those tests.

1. Connect the intended project

Cursor IDE with hosted MCP

For browser sign-in, merge this entry into the project’s .cursor/mcp.json. The workspace name below lets you target this connection even if the PyAI plugin has a separate saved login.
  1. Open the project in Cursor. In Customize → MCPs, open the scope menu and select the workspace. The User filter hides project entries.
  2. Open pyai-workspace, enable its workspace source and select Authenticate under Local.
  3. On the PyAI consent page, choose your organization and project. Check the displayed project ID and environment. Select Read-only access, review the permissions shown, then select Connect to PyAI.
  4. In a new chat in that workspace, send this prompt before asking Cursor to do anything else.
Match the IDs and environment to the project you chose. For a test project, expect env: test and status: active. Compare scopes with the consent preview. The original 19 November login returned three permissions: omni:read for call records, trace:read for compliance findings and amd:read for answering-machine decisions. A client requesting fewer permissions can receive a subset. If the project or permissions differ from what you approved, stop, log out of Local on this workspace entry, authenticate again and repeat whoami. Find an existing summary. For Recap, a fresh read-only login can also include recap:lookup. Check for it in whoami, then ask Cursor to use list_recap_calls to find a completed summary or get_recap if you know its call ID. Recap must be enabled in the selected organization. Its records are shared within that organization. Earlier read-only logins keep their original permissions until you reconnect and approve the updated list. Creating a summary requires recap:create. Generation and configuration require a new approval. Before any billable action, repeat the identity check on the newly authorized connection and confirm the billing terms, effective limits and approved spend. On 19 November 2026, Cursor IDE 3.30.8 on macOS verified the selected active test project with whoami and returned the expected workflow or limitation for all nine discovery prompts. The test stopped before submission. The saved-WAV walkthrough below uses a separate local connection.

Local setup for the saved-WAV walkthrough

Use Node.js 22+ and a private CLI profile. The commands below are for macOS/Linux; the MCP JSON configuration works wherever Node and npx are available. Keep credentials out of prompts, source files and MCP configuration. If you already have an authorized profile, reuse its name in every command and configuration below. Otherwise, sign in and approve the intended organization and project, checking the project ID and environment rather than accepting the selected default:
Inspect the profile without generating audio:
env -u prevents inherited credentials or an endpoint from overriding the selected profile. Login saves a private credential; it does not export an API key into your shell. New sandbox creation is a separate, explicit operation in Getting started, not a fallback after an unexpected project or quota error.

Cursor

Merge this entry into your project’s .cursor/mcp.json, preserving other servers:
The empty environment overrides make the saved profile authoritative. In Cursor’s Customize → MCP view, enable this server and confirm it connects. A configuration file alone does not establish that the server is enabled. If Cursor cannot find npx, use its actual absolute path from command -v npx.

Claude Code

Merge the same JSON into the project-root .mcp.json. Open Claude Code in that project, approve the project MCP configuration when prompted, and use /mcp to check pyai-first-result. Review shell/file permissions needed for the saved-file steps; do not disable the client’s permission protections. For unattended --print / --resume runs, use Claude Code 2.1.326, the version verified to connect before the first tool call. The older 2.1.177 started a resumed turn while MCP was pending and attempted an unavailable tool. Claude Code documents startup waiting with --mcp-config from 2.1.221; pass the explicit config and --strict-mcp-config, and inspect connection status before treating a missing tool as a PyAI service failure. See the Claude Code CLI reference. Keep test guards focused on the authorized project and submission counts. Allow ordinary read-only CLI/file inspection; a blocked setup command is test friction, not evidence of a failed speech or transcription request. This setup launches local stdio. For hosted HTTP with browser OAuth, use the MCP setup guide. A hosted connection and a local CLI profile are separate authorizations: verify both identities before using a CLI fallback. Hosted HTTP does not write a file onto your machine.

2. Run a read-only preflight

Ask the agent:
Confirm speak:synthesize and hear:transcribe access, the intended org_id and project_id, and the allowed spend. A ten-unit cap, credit.gated: false, or a null key budget does not establish that a connection is free. For a zero-spend trial, use a project whose nonbillable status has been confirmed; stop if that cannot be established. See pricing for live usage. Client model subscriptions or credits are separate from PyAI usage. Choose a bound before proceeding: one synthesis, one transcription, this one short sentence, no automatic retries, and no other audio workflows. Setting a budget in a prompt does not create a server-side spending limit.

3. Generate and save once

After approving the preflight and filling in your own project IDs, give the agent:
Local output_path is relative to the MCP server’s working directory. Its parent must already exist and stay within that directory; existing files are rejected before submission. Use the returned absolute path for subsequent commands. The saved-file result contains output_path, bytes and status. Tool success alone does not prove audible playback. With hosted HTTP, download the returned audio_url/resource_link once and record its local path. Links are bearer capabilities, expire after one hour and are subject to connection revocation; keep them private. Local stdio without output_path returns inline audio instead. Host player and download support vary; see the audio result contract.

4. Play and transcribe the saved file

Open the file in a local player. On macOS, with the Mac unlocked:
A successful player exit and a person’s confirmation that the words are audible are separate checks. If playback fails, fix the player/device issue and replay the same file; do not synthesize again. transcribe_audio accepts base64 audio, not a local pathname. For the saved-file path, the tested fallback uploads the file through the CLI without passing its base64 contents through the model:
This command submits audio to PyAI’s cloud. It cannot satisfy an offline, on-device or no-upload requirement. Run it only after the usage authorization above. Expected text: The blue lantern is ready for tomorrow. Report the actual text and any difference; do not replace the observation with the expected answer. Save the transcript alongside the WAV and retain their paths.

Recover without duplicate work

What was actually tested

On November 13, 2026, both macOS clients used local MCP 0.5.11, one synthesis and one CLI transcription each, in the same verified nonbillable test project. Each downloaded a 153,644-byte WAV containing 3.2 seconds of speech. Both transcriptions matched the sentence above, and a person confirmed both recordings were clear and intelligible. PyAI credits were unchanged. Those runs exposed unfinished WAV length fields. MCP 0.5.12 fixed the cause by requesting complete speech output; repaired copies retained identical audio and passed duration/playback checks. SDK/CLI 0.18.0 also shipped the saved-WAV fix. Current setup pins are maintained from the published release manifest. The following results retain the versions actually tested. MCP 0.5.19 setup and discovery were verified on November 17. A separate one-clip Claude Code test saved a valid WAV in 17.9 seconds; human playback and exact transcription passed after a client update and test-guard assistance. That was an assisted recovery, not an uninterrupted benchmark. The historical recordings are not a fresh audio test of those newer versions.

Questions about the result

Is this a fully autonomous, timed benchmark? A later configured-workspace Claude Code 2.1.326 / MCP 0.5.19 run on November 17 saved a valid WAV in 10.620 seconds and completed the two-turn speech/transcription journey in 44.643 seconds, with one synthesis, one transcription, exact text and human playback confirmation. It needed no execution intervention. This excludes cold installation and does not supply comparable Cursor/Codex timings. Earlier assisted and failed attempts above remain part of the evidence. For a future authorized benchmark, record connection start, verified identity, request start, saved file, first playback and transcription completion using one clock, along with permissions, errors and request counts. Can a hosted URL prove that the file played? No. Verify download, file validity, player completion and audible speech separately. Does a test label guarantee zero spend? No. Confirm the actual project and billing treatment before usage, then keep request and input bounds explicit. Host setup references: Cursor MCP and Claude Code MCP. For a full application, continue with the SDK starters.