BYOK — Bring your own AI key
Connect a personal API key from any major AI provider (Anthropic, OpenAI, Gemini, Mistral, Groq, DeepSeek, xAI, or any OpenAI-compatible endpoint) and pick a model. The user's key, the user's quota.
Status: partially live. Connecting a key, picking a provider, and selecting a model are available now in Settings → Connect an AI, and power the AI features shipping today (e.g. Google review summaries, pipeline generation from a description). The broader in-app assistant described below is still on the roadmap.
The BYOK ("bring your own key") integration lets the user paste an AI provider API key into the outsend settings and use an AI assistant directly inside the app — to configure searches, draft filter rules, summarise results, or build pipelines through natural language.
Why BYOK and not a hosted model
- The user's spend stays on the user's account, billed by the AI provider directly.
- No outsend-side mediation: the assistant sees only what the user grants it.
- Provider choice stays with the user: Anthropic, OpenAI, or any compatible endpoint.
Supported providers
The provider and model are chosen in Settings → Connect an AI. Models are detected live from the provider's own API — there is no fixed model list to maintain, and new models appear automatically as the provider releases them.
| Provider | Key format | Notes |
|---|---|---|
| Anthropic (Claude) | sk-ant-… |
Native Messages API |
| OpenAI | sk-… |
Incl. reasoning models (o-series, GPT-5) |
| Google (Gemini) | AIza… |
OpenAI-compatible endpoint |
| Mistral | — | |
| Groq | gsk_… |
|
| DeepSeek | sk-… |
|
| xAI (Grok) | xai-… |
|
| Any OpenAI-compatible endpoint | — | Paste a custom base URL (Together, Perplexity, OpenRouter, local Ollama / vLLM, …) |
The key is stored encrypted at rest (Fernet, server secret), scoped to the user's account, and never sent outside the outsend backend except to the chosen provider. A rough cost estimate is shown before AI actions — it is indicative only (best-effort token counting against known public prices) and may differ from the provider's actual billing. AI spending is also protected by hard caps — per request, per day and per month — that you set in Settings: a request that would exceed a cap is blocked before the provider is called, with email alerts at 80% and when a cap is reached. See AI spending caps.
What the assistant can do
The assistant uses the same outsend API surface documented in API overview. It can:
- Read the user's jobs, pipelines, and veilles
- Start new jobs (with explicit user confirmation for spend)
- Compose pipelines by chaining modules from the registry
- Compute filter rules from natural-language descriptions and preview the result
It cannot:
- Access other users' data
- Modify billing, account settings, or invitation codes
- Run anything outside the user's normal permission scope
Why not just use Claude.ai with outsend as MCP?
Both options will exist:
- BYOK — for users who want the assistant inside outsend.xyz, with the UI rendering search forms and tables natively while the model orchestrates.
- MCP — for users who want to drive outsend from their own Claude.ai or Claude Desktop, with their existing subscription.
The two patterns are complementary, not competing.
What's next
- MCP integration — drive outsend from your own AI client
- llms.txt — point any AI assistant at the docs