What is supported
The support tiers, platforms, models, media, voice, memory search, and coding engines this release supports.
Use this page to check whether something is a supported feature. A command or adapter existing in the code is not, on its own, a supported workflow.
Support tiers
| Tier | What it covers |
|---|---|
| A | The local runtime and CLI, the web app, chat, local coding, browser automation, skills, approvals, memory, the daily brief, and backups. |
| B | Connected accounts with explicit sign-in and health checks: email, calendars, tasks, contacts, and home control. |
| C | Long-tail messaging channels, phone and device apps beyond the Mac, and plugins. Experimental or for advanced users. |
Platforms
| Where Oktapod runs | Support |
|---|---|
| macOS | Apple Silicon only. Mac app, LaunchAgent service. |
| Linux | x64 and ARM64. Setup app (DEB, AppImage), systemd --user. |
| Windows | x64. Setup app, Scheduled Task service. Windows on ARM runs the x64 build through Windows' emulation. |
| Client | Support |
|---|---|
| Web app | Any current browser. Installable as an app. |
| iPhone | iOS 17 or later. In App Store review. Tier C. |
| Android | Android 9 or later, from Google Play. Tier C. |
| Command line | Installed with Oktapod on every platform. |
Meetings
| Source | Support |
|---|---|
| Live transcript (web app) | Browser speech recognition in Chrome, Edge, or Safari. Firefox is not supported. |
| Meeting summary | Off until a model is allowed to read private content; allowing it is an operator step on a source install. |
| Teams transcripts | Work or school Microsoft accounts, for meetings someone transcribed in Teams. |
| Transcript files | VTT and SRT import is an operator tool on a source install. |
Models and media
Oktapod keeps a capability record for known models: whether they accept images, audio, and video, and what native attachment limits they have.
- An attachment adds a requirement when the model is chosen. A capable model can receive the file directly.
- A text-only model gets extracted text, a transcript, or a summary, with citations.
- If no capable path exists, the request fails and says why. Oktapod does not pretend it understood a file.
- Unknown model IDs are allowed, and treated as having no special media support.
Documents: PDF text is extracted directly. Scanned or image-only PDF pages are read by a vision-capable model when one is connected. Audio is transcribed by the configured speech service; video is sampled into frames plus a transcript.
Voice
| Direction | Services |
|---|---|
| Speech to text | OpenAI, xAI, Deepgram, Google, Groq, Whisper (local), Parakeet (local) |
| Text to speech | OpenAI, xAI, Gemini, ElevenLabs, Edge, Kokoro (local), Piper (local) |
Sending synthesized audio to someone is an outbound action and follows the approval rules.
Memory search
Memory uses keyword search plus semantic (vector) search.
- Default: a small local embedding model (English-focused) that runs on your machine. Setup warms it up and checks it before finishing.
- Optional: a multilingual local model, an online embeddings service (OpenAI or any OpenAI-compatible endpoint), or off. Change it in Advanced settings → Memory search; switching rebuilds the index.
If the local embedding worker is unavailable, Oktapod reports degraded search and keeps keyword search working. It never silently switches to a test provider.
Coding engines
- Codex, Claude, and Gemini are fully supported and run through Oktapod's own coding runtime.
- Other engines (such as Cursor, Copilot, OpenCode, or Qwen) need an explicit command and are not supported the same way.
- Azure OpenAI can be a model provider, but it is not a coding engine.