Get started
Build with AI
Give your coding agent the API knowledge and tested tools it needs to build integrations and run reliable evaluations.
Install the Omi API skill
Run one command in your project. The same portable skill supports Claude Code, Codex and other agents that use Agent Skills. It includes API guidance, protocol references and a Python evaluation runner.
Claude Code
npx skills add https://docs.omi.health/skills/omi-api.zip --agent claude-codeUse /omi-api after installation.
Codex
npx skills add https://docs.omi.health/skills/omi-api.zip --agent codexUse $omi-api after installation.
Reload your agent session if the skill does not appear. Installation uses the open-source Skills CLI and does not create keys, change billing or send audio.
For manual installation, extract the omi-api folder into .claude/skills/ or .agents/skills/. Desktop and chat products may require a separate upload; CLI installation does not enable it there automatically. Other agents can read agents.md directly.
Give your agent a task
Use the Omi API skill to evaluate omi-medical-1 on my WAV files through the realtime websocket. Use my existing OMI_API_KEY environment variable. Run a dry run and a small smoke first. Keep failures and report first visible text, confirmed turns and the closing record separately, with the served model and client location.The runner includes paced audio, verified model selection, resumable attempts, per-file results, WER and clearly named latency clocks. It keeps API keys out of reports. Local outputs contain transcripts; keep them private. Live runs use your normal allowance and pricing.
Read the runner instructions and output definitions. Default concurrency is one; increase it within your project's available limits.
Choose the right test
| What you want | What to use |
|---|---|
| Live transcription or voice agent | Realtime websocket; pin Medical or General and verify ready.model.served. |
| A recording with custom vocabulary | Batch omi-medical-1. Compare terms on/off with stored dictionaries disabled, and inspect the audit. |
| Live transcription with custom vocabulary | Not currently supported. Batch vocabulary fields do not apply to the realtime start message. |
| Known participants on separate audio channels | Use the channel identity; enable speaker attribution only when you need it. |
Audio length and delivery are different
Maintained documentation
The skill links back to the current public contract. Give any agent https://docs.omi.health/agents.md for a compact starting point, docs.md for the complete guide, and OpenAPI for the HTTP schema. Package version and hashes let you check which skill you installed.
For a vocabulary experiment, the agent should report prompt coverage and delivered terms separately. For streaming, it should separate mutable previews, immutable confirmed turns and the closing record. A successful request alone does not prove a feature was applied.
Privacy, billing and retry rules remain the same when using a coding agent.