Answers teams need before they pilot, procure, and launch.

An MCP proof of concept can be running in minutes: add the server configuration, set an API key and call an assessment tool. A production integration takes longer because your team still needs to validate representative audio, thresholds, retry behavior, privacy and monitoring.

Cursor, Claude Desktop, Cline, Windsurf, Zed, and any other MCP-compatible client. Also works as a tool inside LangChain, LlamaIndex and the OpenAI Agents SDK via the MCP adapter.

Use the SDK or direct API when your product owns audio capture and the assessment flow. Use MCP when you want compatible agents to discover and call a standardized tool. Use function calling when you already own the agent loop and want direct control over validation, retries and routing.

Mandarin Chinese and English are first-class, both with phoneme-level scoring. Chinese includes dedicated handling for tones, pinyin, neutral tone, erhua and tone sandhi. English includes CEFR-aligned scoring with stress and rhythm diagnostics.

Yes. A WebSocket streaming session accepts mic audio frames and returns scores within a few hundred milliseconds of end-of-speech. File evaluation supports mp3 / wav / m4a / ogg / aac / pcm.

Your integration should treat unusable or incomplete audio as a recording problem, not a pronunciation result. Use structured error and quality signals to request a new recording without penalizing the learner or presenting a misleading score.

Yes. Chivox returns assessment evidence; your product decides thresholds, levels, pass rules, retry logic and how much detail to show. This keeps pedagogy, tone of voice and accessibility choices under your control.

Accuracy depends on the language, task type, speaker group, rubric and recording conditions. Chivox is built on large-scale production assessment, but your pilot should still compare representative user audio with your own expert judgments before you set product thresholds.

The standard streaming assessment path scores audio in memory and returns structured results without retaining an additional audio copy in the scoring runtime or using that audio for model training. Confirm any customer-specific logging, support or deployment requirements during security review.

Yes. Enterprise discussions can cover data flow, retention, authentication, environment separation, deployment, service levels and contractual requirements. Bring your questionnaire and expected regions early so the team can identify any plan-specific constraints before launch.

MCP and function-calling plans use published evaluation-point pricing and include free starting credits. SDK and direct API licensing is quoted separately based on product, volume, deployment and support requirements.

No. Published self-serve MCP and function-calling plans deduct points only for successful evaluations. Rejected audio, errors and failed calls do not reduce the balance.

Bring expected monthly volume, task types, languages, regions, deployment preference, target service levels, security requirements and a small set of representative audio. That is enough to scope a useful pilot and identify procurement questions early.