Official MCP server for Listenhub - The AI Voice for Every Creator. Provides tools for creating podcasts and FlowSpeech episodes with AI-powered text-to-speech generation.
The server defines 8 tools with generally good naming (verb_noun convention) and comprehensive descriptions. However, there are significant gaps in output schema documentation, parameter descriptions could be more precise, and error handling lacks recovery guidance. Most tools have solid descriptions (100-300 chars, within baseline range), but schema documentation for return types is either missing or not visible in the provided source. Tool parameters are well-typed with enums where appropriate, but several lack usage examples or constraints. The two-stage podcast creation pattern (create_podcast vs create_podcast_text_only) is well-designed but creates potential user confusion. No security concerns detected, API keys are injected server-side, not exposed as parameters.
Create a FlowSpeech episode by converting text or URL content to speech. Supports smart mode (AI-enhanced, fixes grammar) and direct mode (no modifications). This tool will automatically poll until generation is complete.
Create a podcast episode with full generation (text + audio). Supports single-speaker (solo) or dual-speaker (dialogue) formats with 1-2 speakers (can use speaker names or IDs). Choose from 3 generation modes: quick (3-5 min podcast), deep (8-15 min podcast), or debate (5-10 min podcast). Accepts text or URL sources. This tool will automatically poll until generation is complete (may take several minutes).
Create podcast episode with text content only (no audio generation). Supports single-speaker (solo) or dual-speaker (dialogue) formats with 1-2 speakers (can use speaker names or IDs). Choose from 3 generation modes: quick (3-5 min podcast), deep (8-15 min podcast), or debate (5-10 min podcast). This is the first stage of two-stage generation. After text generation completes, you can review the scripts. To use modified scripts, call generate_podcast_audio with customScripts parameter to override the generated scripts when generating audio.
Generate audio for a podcast episode that was previously created with text-only generation. Optionally accepts customScripts to override the original generated scripts. This allows you to modify the podcast scripts before audio generation.
Output schemas not documented. While tools like create_podcast and create_flowspeech have detailed input schemas, their return types are not specified in the source code provided. The LLM cannot know what fields to expect (e.g., does create_podcast return episodeId, audioUrl, scriptText?). This breaks tool chaining and forces discovery calls.
Parameter 'sources' in create_podcast and create_podcast_text_only is marked optional but documentation does not state: (1) what happens if both 'query' and 'sources' are omitted? (2) how many sources are allowed? (3) does URL validation happen server-side? This ambiguity may lead LLMs to pass invalid combinations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 65 | 2026-07-28+ | v2 |
| 2026-03-09 | C | 61 | - | v1 |
Query detailed information of a FlowSpeech episode, including generation status, audio URLs, scripts, outline, and metadata. Does not poll - returns current status immediately.
Query detailed information of a podcast episode, including generation status, audio URLs, scripts, outline, and metadata. Does not poll - returns current status immediately.
Get list of available published speakers for podcast generation. Supports filtering by language code (e.g. "zh", "en"). Returns speaker ID, name, language, gender and demo audio URL. Defaults to Chinese speakers if no language specified.
Get current user subscription information, including subscription status, credit usage, plan details, and renewal status.
create_podcast description states 'This tool will automatically poll until generation is complete (may take several minutes)' but does not specify a timeout, max retry count, or how the LLM is notified of polling completion. An LLM cannot distinguish a hung call from normal long latency. No polling strategy or timeout bounds are documented.
Parameter descriptions for 'speakers' and 'speakerIds' mention 'Use speaker names from get_speakers tool output (the "name" field, not speakerId)' but also accept 'Full speaker IDs also supported.' This dual acceptance is good for UX but the parameter description should explicitly state 'Accepts either speaker name (from get_speakers output) OR full 24-character speaker ID' to remove ambiguity.
get_podcast_status and get_flowspeech_status require exact 24-character episode IDs but do not provide error guidance if the ID is malformed or does not exist. Description states 'IMPORTANT: Must be exactly 24 characters' but does not explain what happens if validation fails or how the LLM should recover. Error handling is missing.
create_podcast_text_only and generate_podcast_audio are designed as a two-stage workflow, but the connection between them is not made explicit in the descriptions. An LLM may not understand that generate_podcast_audio is only valid for episodes created via create_podcast_text_only, not via create_podcast. Dependencies are undocumented.
No explicit error handling or recovery guidance documented across all tools. If an API call fails (network, rate limit, invalid input), what should the LLM do? Should it retry? Ask the user? There are no error codes, messages, or categorization (retryable vs. fatal) visible in the schemas.
Parameter 'mode' in create_podcast and create_flowspeech is well-described with time estimates (quick 3-5 min, deep 8-15 min) but does not clarify cost differences or token/credit implications. Users often ask 'which mode should I use?', the tool should provide guidance on when to use quick vs. deep, not just time ranges.