Skip to main content
Glama

Kubeez

generate_dialogue

PRE-FETCH (mandatory before polling the result): resources/read kubeez://skills/polling.

Scope: generate:speech. Generate text-to-speech audio. FOUR providers, select with provider: elevenlabs (default, ElevenLabs v3), google / gemini (Google Gemini 3.1 Flash TTS), gemini-3.8 (Gemini 3.8 Flash TTS: TWO-SPEAKER DIALOGUE with a delivery style per line), or seed (Seed 1.0 Audio: prompt-driven, supports image/audio references). ROUTING (the parameter default stays elevenlabs for backward compatibility; it is NOT the recommended pick): provider='gemini-3.8' is the recommended default for narration, bulk TTS, two-person dialogue / podcasts and simple expressive reads. provider='google' when text carries free-form [square] tags, when you need explicit language_code control, or for a long single take. provider='seed' for a scene with sound effects or music, and for voice cloning. provider='elevenlabs' only when the user names an ElevenLabs voice or wants its stability controls. ElevenLabs strips tags and ships flat audio, so leaving the default on an expressive or tagged ask is the #1 mistake. Safety net: tagged text sent with the ElevenLabs default (and no explicit ElevenLabs voice) is auto-routed to Gemini, but set provider='google' yourself rather than relying on it. NOT music generation: do not use prompt, instrumental, or the Suno V6 model fields — use generate_music for songs. Required: text — the literal words and sounds to be SPOKEN, read VERBATIM. Two mistakes to avoid on EITHER provider: (1) a DESCRIPTION ('sensual vocalizations of a woman pleasuring herself' / 'soft panting') — the model just SAYS those words; (2) a bare LIST of bracketed tags ('[moaning softly], [gasping gently], [light panting]') — nothing to actually say. [bracket] performance cues are a GEMINI-ONLY feature (provider='google'): INLINE guides placed right before the words/sounds they shape, woven INTO real spoken content — e.g. text = '[about to have an orgasm] Ohhhh im going to [slight cry] cuummm, i cannot keep it anymore'. ELEVENLABS STRIPS every [tag] before synthesis (never performed), so ANY expressive / tagged / wordless-vocalization (moans, sighs, panting) ask MUST use provider='google'; on ElevenLabs text is just the plain spoken words (bare phonetic sounds like 'Mmm... ahh...' at most, no tags). For a Gemini wordless vocalization, write the sounds PHONETICALLY ('Mmm... ahh... ohhh') with inline cues between them ([moaning], [gasp], [breathing heavily]). Character/tone go in the voice choice; keep style_prompt minimal (default 'Say the following.', see below) — NOT a scene description. If you have only a vibe and no words/sounds to perform, ASK the user what should be said. voice defaults to Rachel (ElevenLabs) / Kore (Google) if omitted. Limits below are rejected with 400 and NO credit charge.

ELEVENLABS (provider='elevenlabs'): • text 5-5000 characters after audio tags are stripped. • Audio tags ([HEY], [laughs], [whispers], …) in text/previous_text/next_text are STRIPPED server-side — NOT performed. • voice: one of 26 — Rachel, Drew, Clyde, Paul, Aria, Domi, Dave, Roger, Fin, Sarah, James, Jane, Juniper, Arabella, Hope, Bradford, Reginald, Gaming, Austin, Kuon, Blondie, Priyanka, Alexandra, Monika, Mark, Grimblewood. • language_code: one of 29 ISO codes (empirically verified — upstream rejects v2's hu/no/vi/he/fa): ar, bg, cs, da, de, el, en, es, fi, fil, fr, hi, hr, id, it, ja, ko, ms, nl, pl, pt, ro, ru, sk, sv, ta, tr, uk, zh. Pass auto to default to en. • Tuning: stability (0-1, default 0.5), similarity_boost (0-1, default 0.75), style (0-1, default 0), speed (0.7-1.2, default 1.0), previous_text/next_text (context, tags stripped). Call get_limits_for_model('text-to-dialogue-v3') to confirm the current lists.

GOOGLE / GEMINI (provider='google' or 'gemini'): • text ≤4000 BYTES (UTF-8). Inline expressive tags ([sigh], [laughing], [whispering], [shouting], [extremely fast], [long pause]) are PRESERVED and performed (unlike ElevenLabs, which strips them). • voice: one of 30 — Zephyr, Puck, Charon, Kore, Fenrir, Leda, Orus, Aoede, Callirrhoe, Autonoe, Enceladus, Iapetus, Umbriel, Algenib, Despina, Erinome, Laomedeia, Achernar, Algieba, Schedar, Gacrux, Pulcherrima, Achird, Zubenelgenubi, Vindemiatrix, Sadachbia, Sadaltager, Sulafat, Alnilam, Rasalgethi. • style_prompt (Google only): a SHORT global delivery note — overall voice character / tone for a SPOKEN-narration take (e.g. 'Speak like a theatrical vampire host, slow and velvety'), ≤4000 bytes, default 'Say the following.'. Do NOT describe the SCENE or the SOUNDS here (that reads like the thing to say); per-moment performance lives in the inline [cues] woven into text. For wordless vocalizations, leave it at the default and let the inline cues + phonetic sounds carry it. text + style_prompt must be ≤8000 bytes combined. • language_code: a BCP-47 code (e.g. en-US, es-ES, ro-RO) or auto to auto-detect. ElevenLabs tuning params (stability/similarity_boost/style/speed/previous_text/next_text) are ignored. Call get_limits_for_model('text-to-dialogue-gemini') to confirm the current lists.

GEMINI 3.8 FLASH (provider='gemini-3.8'): the provider for a CONVERSATION between two voices, or a single voice whose delivery changes line by line. • Simple: text + voice (+ style_prompt = that line's delivery). • Dialogue: speakers = 1-2 items [{"voice": "Fola"}, {"voice": "Zephyr"}] (Speaker 1, Speaker 2 in order) and dialogue_turns = ordered [{"speaker": "Speaker 1", "text": "...", "style": "slow and dry"}, ...] (≤50 turns; style optional, ≤500 chars each, ≤2000 total). text is ignored when dialogue_turns is set. • filler_words=true adds natural um/uh hesitations; only with two speakers. • Tags use ANGLE brackets inline in a turn's text: , , , , , , , , , (40 official; [square] tags are NOT performed here). • voice: 70 voices grouped by use case (default Fola), e.g. Podcast: Alnilam, Brio, Jori, Laomedeia, Rami, Veda; Storyteller: Achird, Enzo, Riko, Sami, Zali, Zubenelgenubi; Commercial: Koda, Kore, Nika, Tavi, Zeno. Full list: get_limits_for_model('gemini-3-8-flash-tts'). • No language_code: the model reads the language from the text. • Limit: ≤5000 characters across all turns, INCLUDED (tags are billed like text). ElevenLabs tuning params are ignored.

SEED 1.0 AUDIO (provider='seed'): PRE-FETCH resources/read kubeez://skills/seed-audio-prompting (or get_skill('seed-audio-prompting')) BEFORE generating — it's a PROMPT-DRIVEN generative model (voices + sound effects + background music from one prose 'director's script'), and a bare line gives bare output. text is a free-form prompt (≤2048 chars), NOT a fixed voice reading verbatim; there is NO voice param and [tags]/ElevenLabs tuning do NOT apply — describe each voice (gender/age/accent/timbre/emotion) in the prose. Optional references (image and audio CANNOT be mixed): one IMAGE — image_url or base64 image_data — to guide voice/mood; OR up to 3 AUDIO references via audio_references (a list of {speaker | audio_url | audio_data}, each clip ≤30 s) for voice cloning. Tag each cloned voice IN THE PROMPT by upload order: <<TGT_SPK1>> / <<TGT_SPK2>> / <<TGT_SPK3>> is canonical (e.g. 'Marcus (..., the actor is <<TGT_SPK1>>) says: "…"'); @Voice1 / @Audio1 / @Speaker1 are also accepted and auto-normalized to <<TGT_SPKn>> server-side, so either form binds. Optional audio_config {format, sample_rate, speech_rate, loudness_rate, pitch_rate}. Billed by OUTPUT SECONDS (capped at 120 s) — read the per-second rate from get_models / estimate, don't assume.

Billing (elevenlabs/google/gemini-3.8): PER 1000 characters (decimal ≤0.3 rounds down, >0.3 rounds up; minimum 1 credit) — read the per-1000-char rate from get_models / estimate, don't assume. Requires scope generate:speech or generate:music. Returns generation_id; poll get_status(id) until status is completed and output URLs are non-null (CDN-only). Outputs appear in media_outputs with media_type=audio.

Tool call template

This tool declares no required parameters. Add optional parameters from the input schema as needed.

{
  "id": 1,
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "arguments": {},
    "name": "generate_dialogue"
  }
}

Input Schema

NameRequiredDescriptionDefault
textNo
speedNo
styleNo
voiceNo
providerNoelevenlabs
speakersNo
image_urlNo
next_textNo
stabilityNo
image_dataNo
audio_configNo
filler_wordsNo
style_promptNo
language_codeNo
previous_textNo
dialogue_turnsNo
audio_referencesNo
similarity_boostNo

Output Schema

NameRequiredDescriptionDefault

No arguments

Schema Changelog

Changes observed during successful MCP inspections.

  1. First observed

TDQS

A4.8/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Beyond the annotations (readOnlyHint=false, idempotentHint=false, openWorldHint=true, destructiveHint=false), the description discloses the async lifecycle (returns generation_id, poll get_status until completed, CDN-only URLs in media_outputs), billing mechanics per 1000 chars / per output second, 400-rejection limits with no credit charge, required scope, and an auto-rerouting safety net. These are exactly the behavioral facts annotations cannot carry.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness3/5

Is the description appropriately sized, front-loaded, and free of redundancy?

It is front-loaded well — prefetch, scope, then routing — and uses provider headers for scannability. But the ElevenLabs-strips-tags fact is restated at least four separate times ('strips tags and ships flat audio', 'STRIPS every [tag]', 'Audio tags ... are STRIPPED server-side', 'unlike ElevenLabs, which strips them'), and the description is long enough that an agent may truncate before reaching the seed billing rules.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For an 18-parameter, four-provider, asynchronous generation tool with 0% schema coverage, the description covers inputs, limits, output lifecycle, and cost. The output schema exists so return-value explanation is not required, and it still notes the shape (generation_id, media_outputs with media_type=audio).

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters5/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 0%, so the description carries the full burden and does — all 18 parameters get meaning: text vs. dialogue_turns precedence ('text is ignored when dialogue_turns is set'), speakers ordering, filler_words requiring two speakers, audio_config fields, image/audio reference exclusivity, and the ElevenLabs-only tuning params that are ignored on other providers. Voice enumeration and per-provider semantics are added where the schema has none.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The opening states a specific verb and resource ('Generate text-to-speech audio') with the scope declared ('generate:speech'), and immediately separates it from the nearest sibling: 'NOT music generation ... use generate_music for songs.' Provider names and routing make it unmistakable which ask belongs here versus generate_music.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines5/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

Routing is explicit and normative: gemini-3.8 is the recommended default for narration/bulk/dialogue, google for [square]-tagged text or language_code control, seed for sound effects/music/cloning, elevenlabs only when a user names a voice. It also gives an exclusion rule ('If you have only a vibe and no words/sounds to perform, ASK the user') and warns that the default is not the recommended pick — far beyond typical when-to-use guidance.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Try in Browser

Glama MCP Gateway

Add one secure layer between your agents and this server.

Resources