# LatenTX Realtime Voice models for complex Agent tasks, Skills and MCP tools. ## Start here for coding agents Read the LatenTX Agent Skill: https://api.latenth.com/skills/latentx/SKILL.md One developer Skill covers API integration, business Prompt/Skill/MCP configuration, IM apps and browser voice apps. Load only the task-specific references linked from SKILL.md. Installation guide: https://api.latenth.com/docs/ai-skills.html Download https://api.latenth.com/skills/install.py, review it, then run `python3 install.py latentx --agent codex` or `python3 install.py latentx --agent claude`. The installer checks the published SHA-256 manifest and preserves existing installations. Reading or installing the Skill does not call a paid model; real API calls are metered. Keep API keys server-side. ## Realtime API - Base URL: https://api.latenth.com - WebSocket: wss://api.latenth.com/v1/realtime - Examples (discover accessible IDs with GET /v1/models): latentx-realtime-2.5-max, latentx-realtime-2.5-plus, latentx-realtime-2.5-mini, latentx-realtime-2.5-multilingual-max, latentx-realtime-2.5-multilingual-mini - Multilingual models default to intl-female. Read model-specific input_languages, output_languages, voices and price from GET /voice/v1/catalog; never infer language codes or reuse another model’s voice. - [Integration guide](https://api.latenth.com/docs/quickstart.html) - [Realtime API reference](https://api.latenth.com/docs/realtime-api.html) - [Voice management API reference](https://api.latenth.com/docs/api-reference.html) - [Custom models and voices](https://api.latenth.com/docs/custom-models.html) - [Web and phone integration](https://api.latenth.com/docs/phone-integration.html) - [Console](https://api.latenth.com/voice) - [Realtime WebSocket test page](https://api.latenth.com/realtime/) - [WebRTC voice playground](https://api.latenth.com/playground) - [GPT text models and Responses API](https://api.latenth.com/docs/gpt-models.html) Uses OpenAI Realtime GA-style WebSocket events. See the guide for supported features and billing semantics. Start with an API key, model and instructions; no Agent or model binding is required. Function tools and remote MCP use standard tools configuration. Optional inline Skills and saved Skill/Knowledge/MCP references use the documented session.latentx extension, which is not OpenAI-native. The same key authorizes tenant-scoped management APIs. Full API guide: https://api.latenth.com/docs/realtime-api.html Realtime usage: the default response mode reports usage in response.done.response.usage. Opt in at connection time with X-LatenTX-Usage-Mode: transcription or ?usage_mode=transcription. Confirm session.latentx.usage_mode on session.created. In this mode, count duration usage on conversation.item.input_audio_transcription.completed even without a response; do not count ASR again in response usage. See https://api.latenth.com/docs/realtime-api.html#billing. Read current model capabilities and sales prices from GET /voice/v1/catalog; do not hard-code rates. - [Custom model publishing and voices](https://api.latenth.com/docs/custom-models.html) Tenant-owned custom models use the same Realtime protocol with their published custom-...-vN model ID. Discover allowed models with GET /v1/models. Custom model config.skills accepts only saved {id,version,preload} references: save the business Skill first with PUT /voice/v1/skills/{id}/versions/{version}; inline markdown is for session.latentx.skills and fails custom-model validate with 400. Every session carries a fixed session time reference (start time, time zone, weekday) after the instructions. Knowledge uses session.latentx.knowledge_refs or custom model config.knowledge_refs and server-executed search_knowledge over attached resources only. The tenant-scoped management API is rooted at https://api.latenth.com/voice/v1 and covers catalog, voices, custom models, Skills, MCP resources, temporary web credentials, phone bindings, conversations, admin logs and usage. Voice selection uses session.audio.output.voice; catalog: GET /voice/v1/voices. Use voices and default_voice from the selected catalog model; voice locks after first audio. Web/phone credentials are issued server-side; the browser SDK is at /sdk/latentx-voice.js. - [Browser voice OpenAPI (credentials and conversations)](https://api.latenth.com/docs/webrtc.openapi.json) SDK supports inputStream, onPeerConnection, getStats(), and once-per-call onEnded. ASR silence_duration_ms is separate from call idle_timeout_ms. WebRTC uses SIP signaling and is not OpenAI native WebRTC. Browser SDK and Console Playground calls are metered and settled through billing sessions. Inspect the returned billing session and final settlement state; do not infer settlement from a successful call or hangup. Phone binding alone does not establish phone billing availability; verify that channel separately. ## LatenTX Agent Skill Install one developer Skill, `latentx`, in Codex or Claude. Read its entrypoint, then only the references needed for the task. This installed Skill is distinct from runtime business Skills configured on voice sessions. - [Agent setup, installation and examples](https://api.latenth.com/docs/ai-skills.html) - [SKILL.md — the single entrypoint](https://api.latenth.com/skills/latentx/SKILL.md) - [Complete Skill ZIP](https://api.latenth.com/skills/latentx.zip) - [Installer — review before running](https://api.latenth.com/skills/install.py) - [Integrity manifest](https://api.latenth.com/skills/manifest.json) - [Studio Partner — customer-service reference project](https://github.com/aibpo-labs/studio-partner) The Skill routes API integration/troubleshooting to api-integration.md, business Prompt/Skill/MCP design to agent-design.md, text chat to im-app.md, and browser microphone applications to voice-app.md under its references directory. Do not load every module by default. For an existing application, use the official Skill and preserve its stack, authentication and business structure; the reference project is optional. For a complete reference workspace, use Studio Partner and follow its README and project-local Skills for setup, supported capabilities and extension points. H5 demo interactions are distinct from real IM and browser voice calls, which require an authorized platform key and are metered. Design requires separate service configuration as described by the project. Installation of the official Skill does not download or start the project. LatenTX public API contracts remain defined by these official docs; application backend routes are not platform APIs. Custom-model configuration selects base_model and that model's catalog languages and voices. Keep permanent keys server-side. ## Protocol reference Two paths: predefined LatenTX model + session.update instructions/tools for the current connection; or an existing published custom model ID carrying its configuration. Neither requires independent Agent registration. Existing custom-model creation, publication, Playground and phone behavior is unchanged. - [Realtime reference: events, fields, examples, sequences and Schema](https://api.latenth.com/docs/realtime-events.html) - [Realtime event JSON Schema](https://api.latenth.com/docs/realtime.schema.json) - [Admin API reference: endpoints, fields, examples and OpenAPI](https://api.latenth.com/docs/admin-api.html) - [Admin OpenAPI 3.1](https://api.latenth.com/docs/admin.openapi.json) - [Knowledge configuration Skill reference](https://api.latenth.com/skills/latentx/references/knowledge.md) - [Management workflow Skill reference](https://api.latenth.com/skills/latentx/references/management.md) The prompt field is instructions, not instruct. Function and MCP definitions both use tools. Response overrides affect one response only; session changes do not persist to other connections. Always require response.done.response.status=completed and separately track pending MCP calls. ASR/TTS language choices come from the selected catalog model and are configured in custom models or web/phone requests, not arbitrary session.audio language fields.