Bring Your Own AI Provider (BYOAI)

Crow's built-in AI Chat lets you chat with any AI provider directly from the Crow's Nest dashboard. The AI has full access to your Crow tools — memory, projects, blog, storage, and sharing — so you can manage your data through natural conversation.

How It Works

Crow's MCP servers are tool providers. When you configure an AI provider, the Crow gateway acts as a bridge: it sends your messages to the AI, and when the AI wants to use Crow tools (search memories, create blog posts, etc.), the gateway executes those tool calls locally and feeds the results back.

This means:

• Your data stays on your machine
• The AI provider only sees the conversation and tool results
• You can switch providers anytime by changing a few env vars
• Works with free/cheap options like Ollama (fully local) or OpenRouter

Quick Setup

From the Crow's Nest (recommended)

1. Open the Crow's Nest → Settings
2. Find the AI Provider section
3. Select your provider from the dropdown
4. Enter your API key (not needed for Ollama)
5. Click Save, then Test Connection
6. Go to Messages → the AI Chat tab is now active

From .env

Add these to your .env file (or ~/.crow/.env):

AI_PROVIDER=openai
AI_API_KEY=sk-...
AI_MODEL=gpt-4o

No gateway restart needed — the config is hot-reloaded.

Supported Providers

Provider	AI_PROVIDER	Default Model	API Key Required	Notes
OpenAI	openai	gpt-4o	Yes	Get key
Anthropic	anthropic	claude-sonnet-4-20250514	Yes	Get key
Google Gemini	google	gemini-2.5-flash	Yes	Get key
Ollama	ollama	llama3.1	No	Fully local, no API key needed
OpenRouter	openrouter	openai/gpt-4o	Yes	Get key — access 100+ models

OpenAI

AI_PROVIDER=openai
AI_API_KEY=sk-...
AI_MODEL=gpt-4o

Works with GPT-4o, GPT-4o-mini, o1, and any model available on the OpenAI API.

Anthropic

AI_PROVIDER=anthropic
AI_API_KEY=sk-ant-...
AI_MODEL=claude-sonnet-4-20250514

Works with Claude Opus, Sonnet, and Haiku models.

Google Gemini

AI_PROVIDER=google
AI_API_KEY=AIza...
AI_MODEL=gemini-2.5-flash

Uses the Gemini REST API. Works with Gemini 2.5 Flash, Gemini 2.5 Pro, and other available models. The free tier is generous for personal use.

Ollama (Local)

AI_PROVIDER=ollama
AI_MODEL=llama3.1
AI_BASE_URL=http://localhost:11434

Runs entirely on your machine — no API key, no data leaves your network. Install Ollama from ollama.com or use the Crow Ollama add-on (crow bundle install ollama).

Tool Calling with Ollama

Most local models have limited or no function/tool calling support. For best results with Crow tools, use models that support function calling: llama3.1, mistral-nemo, qwen2.5. Without tool support, the chat works but cannot access your Crow data.

OpenRouter

AI_PROVIDER=openrouter
AI_API_KEY=sk-or-...
AI_MODEL=openai/gpt-4o

OpenRouter gives you access to 100+ models from multiple providers through a single API key. Great for trying different models without signing up for each provider separately. Many models have free tiers.

Custom OpenAI-Compatible Endpoint

Any API that implements the OpenAI Chat Completions format works with the openai provider and a custom base URL:

AI_PROVIDER=openai
AI_API_KEY=your-key
AI_MODEL=your-model
AI_BASE_URL=https://your-endpoint.com/v1

This works with vLLM, LM Studio, text-generation-webui, and other OpenAI-compatible servers.

Environment Variables

Variable	Required	Description
AI_PROVIDER	Yes	Provider name: openai, anthropic, google, ollama, openrouter
AI_API_KEY	Depends	API key (not needed for Ollama)
AI_MODEL	No	Model name (uses provider default if blank)
AI_BASE_URL	No	Custom API endpoint (for Ollama, OpenRouter, or self-hosted)

Using AI Chat

Once configured, open Messages in the Crow's Nest. The AI Chat tab appears with:

• Conversation sidebar — Create, switch between, and delete conversations
• Chat area — Send messages, see streaming responses
• Tool calls — Expandable cards showing when the AI uses your Crow tools
• Cancel — Stop a generation in progress

The AI sees your Crow tools as 7 category tools (crow_memory, crow_projects, crow_blog, crow_sharing, crow_storage, crow_tools, crow_discover). It can:

• Recall your memories and store new ones
• Search and manage research projects
• Create and publish blog posts
• Upload and manage files
• Send messages to contacts
• Discover available tools and their schemas

Conversation Context

Each conversation sends the system prompt (generated from your crow.md context) plus the last 20 messages to the AI. Tool results are truncated to 2000 characters to prevent context overflow. The AI can make up to 10 tool call rounds per message.

Token Tracking

Total tokens are tracked per conversation and displayed in the sidebar. This helps you monitor API usage costs.

AI Chat vs External Platforms

You don't have to choose — they work together:

Feature	AI Chat (BYOAI)	External Platforms (Claude.ai, ChatGPT, etc.)
Setup	Configure API key in Settings	Install Crow MCP servers on the platform
Interface	Crow's Nest dashboard	Platform's native UI
AI provider	Your choice (any supported provider)	Platform's built-in AI
Tool access	Full (all Crow tools via gateway)	Full (all Crow tools via MCP)
Data sharing	Same database — both see the same memories	Same database
Best for	Quick interactions from the dashboard, free/cheap AI	Deep work, platform-specific features

Security

• API keys are stored in plaintext in your .env file on your device
• The Crow's Nest is private by default (local network / Tailscale only)
• Chat conversations are stored in your local SQLite database
• Messages are sent to your chosen AI provider's API — they leave your machine
• For fully local operation, use Ollama — nothing leaves your network

Semantic Search (Opt-in)

When an AI provider with embedding support is configured, Crow can enhance memory search with semantic search — finding memories based on meaning, not just keywords.

Requirements

• An AI provider that supports embeddings (OpenAI, Ollama with nomic-embed-text, or Google)
• The sqlite-vec SQLite extension installed on your system
• Both are optional — Crow falls back to full-text search (FTS5) when either is missing

How it works

1. When you store a memory, Crow generates an embedding vector from the content
2. The vector is stored in a memory_embeddings virtual table (powered by sqlite-vec)
3. When you search with semantic: true, Crow compares your query's embedding against stored vectors
4. Results from semantic and keyword search are merged for the best of both approaches

Installing sqlite-vec

On Debian/Ubuntu:

sudo apt install sqlite3-vec

Or install from source: sqlite-vec on GitHub

After installing, run npm run init-db to create the vector table.

LocalAI Bundle

For fully local AI (including embeddings), install the LocalAI bundle:

crow bundle install localai
crow bundle start localai

Then configure Crow to use it:

AI_PROVIDER=openai
AI_BASE_URL=http://localhost:8080/v1
AI_MODEL=gpt-3.5-turbo

LocalAI provides an OpenAI-compatible API running entirely on your hardware — no data leaves your network.

Troubleshooting

"No AI provider configured"

Set AI_PROVIDER in Settings or .env. At minimum you need the provider name.

"API key is invalid (401)"

Double-check your AI_API_KEY. For Anthropic, keys start with sk-ant-. For OpenAI, sk-. For Google, AIza.

"Model not found (404)"

The model name is provider-specific. Check the provider's docs for available models. For Ollama, run ollama pull <model> first.

"Rate limited"

The provider is throttling requests. Wait a moment and try again, or upgrade your API plan.

Tool calls not working

Some models (especially small local models via Ollama) don't support function/tool calling. Try a model that explicitly supports it: llama3.1, gpt-4o, claude-sonnet-4-20250514, gemini-2.5-flash.

Chat input not responding

Check the browser console for errors. The chat uses Server-Sent Events (SSE) for streaming — ensure your network/proxy doesn't buffer or terminate SSE connections.