Everything VibeCody can talk to that is not VibeCody, what each one needs from you, and — the part that usually goes unsaid — whether you need it at all.
The honest summary: you need one AI provider and nothing else. Every other service on this page is optional, and a VibeCody with none of them configured is a complete coding assistant. Nothing here is required to install, build, run the daemon, edit code, run the agent loop, or use the 1,144 skills.
Where credentials live
Every key on this page goes into an encrypted store, never a plaintext file:
| Scope | Store | Path |
|---|---|---|
| Yours, all projects | ProfileStore |
~/.vibecli/profile_settings.db |
| One project’s secrets | WorkspaceStore |
<workspace>/.vibecli/workspace.db |
Set them in Settings → API Keys / Integrations in any desktop shell, or with
vibecli set-key <provider> <value>. They are never written to config.toml,
and an environment variable is a fallback, not the intended path. See
Configuration.
AI providers — pick at least one
25 providers are selectable. All of them stream. The two local ones need no key and no account.
| Category | Providers |
|---|---|
| Local, no key | ollama, lmstudio, vllm, vibecli-mistralrs (in-process) |
| Frontier | claude, openai, gemini, grok, mistral, deepseek, zhipu, minimax |
| Fast inference | groq, cerebras, together, fireworks, sambanova, perplexity |
| Cloud platform | azure_openai, bedrock, copilot, vercel_ai, openrouter |
| Specialist | poolside, claude-code (local CLI passthrough) |
The canonical list is vibecoder/src/hooks/useModelRegistry.ts. Per-provider
setup, models and pricing notes: Providers.
Fully offline is a supported configuration.
ollamaplus a downloaded model means no account, no key, and nothing leaving the machine. See Sizing for what your hardware can run.
Voice
| Service | Needed for | Key |
|---|---|---|
| whisper.cpp (local) | Speech-to-text, on-device | None — a model download |
| Groq Whisper | Speech-to-text fallback | Groq API key |
| Platform TTS | Spoken replies | None. Built into macOS / Windows / Linux |
| Kokoro-82M via MLX | Neural spoken replies | None — make voice-kokoro. Apple Silicon |
| ElevenLabs | Alternative TTS | elevenlabs_api_key + elevenlabs_voice_id |
Local whisper is tried first when a model is downloaded; the cloud path is the fallback, not the default. Voice · Sizing
Productivity integrations
Configured under Settings → Integrations. Each is independent; none is required.
| Category | Services | What you supply |
|---|---|---|
| Gmail, Outlook | OAuth client ID + secret, or access/refresh tokens | |
| Calendar | Google Calendar, Outlook Calendar | Access token |
| Project tools | Linear, Notion, Todoist, Jira | API key; Jira also needs instance URL + email |
| Search & web | Tavily, Brave Search | API key |
| Smart home | Home Assistant | Instance URL + long-lived access token |
| Infrastructure | GitHub, OpenSandbox, any container registry | Token; registry needs URL + user + password |
These are also exposed as MCP tools, so a model can act on them directly.
Messaging gateways
vibecli --gateway <platform> runs the agent as a 24/7 bot. 25 adapters ship
(vibecli/vibecli-cli/src/gateway.rs):
Telegram · Discord · Slack · Signal · Matrix · Twilio SMS · WhatsApp · iMessage (macOS) · Microsoft Teams · Google Chat · Mattermost · IRC · LINE · Twitch · Nextcloud Talk · Nostr · Feishu · DingTalk · QQ · WeCom · Zalo · BlueBubbles · Synology Chat · Tlon · Web chat
Each takes that platform’s own bot token. A message becomes an agent task and the result comes back as a reply.
Sign-in providers
For signing in to VibeCody itself, not for AI: Google, GitHub, GitLab, Bitbucket, Microsoft, Apple (Settings → OAuth Login). Optional — VibeCody does not require an account.
Connectivity
How phones, watches and remote clients find a daemon. Clients race all reachable paths rather than being configured with one.
| Transport | Needs | Use |
|---|---|---|
| mDNS | Nothing | Same LAN. The zero-config path |
| Tailscale | A tailnet | Off-LAN, private, no port forwarding |
| ngrok | An ngrok account | A public URL, quickly |
| Phone relay | Nothing | The phone relays for the watch |
Details and the ordering: Connectivity.
Sandboxing
Optional isolation for agent-run commands: Docker, Podman, and OpenSandbox (hosted, needs an API key), behind one runtime trait.
Model and asset sources
| Source | Pulled for | Credential |
|---|---|---|
| Ollama registry | Local LLMs and embedding models | None |
| Hugging Face | mistralrs models, whisper weights |
None, except gated repos |
Gated models such as meta-llama/* need a Hugging Face token —
vibecli set-key huggingface hf_..., or HF_TOKEN. Without one the daemon says
so at startup and falls back to an ungated default rather than failing silently.
What none of this is required for
Installing · building · running the daemon · the desktop shells · editing · git · the agent loop with a local model · skills · code review · the code index with a local embedding model · voice with the platform engine.
The only hard requirement is something that can answer a prompt — one cloud key, or Ollama on the machine.