Configure a Model Provider
Manasvi uses a language model to generate responses and decide when to use tools. This page explains how to connect it to different providers.
Available providers
| Mode | Provider | Requires | Best for |
|---|---|---|---|
mock | Built-in test adapter | Nothing | Development, testing |
deepseek | DeepSeek API | API key | Default cloud provider |
ollama | Ollama (local) | Ollama installed | Privacy, local inference, no API costs |
openai | OpenAI API | API key | Cloud-based models (GPT-4o, etc.) |
claude | Anthropic Claude | API key | Cloud-based Claude models |
For a dedicated Claude walkthrough, see Connect Claude.
Using the CLI (recommended)
The easiest way to switch providers is through the CLI:
pnpm manasvi models list
Shows your configured providers and which one is active.
pnpm manasvi models add ollama
# or
pnpm manasvi models add deepseek
# or
pnpm manasvi models add openai
# or
pnpm manasvi models add claude
Walks you through the configuration interactively and writes the settings to .env.local.
pnpm manasvi models test
Sends a test request to verify connectivity.
pnpm manasvi models use ollama
# or
pnpm manasvi models use deepseek
# or
pnpm manasvi models use openai
# or
pnpm manasvi models use claude
# or
pnpm manasvi models use mock
Switches the active provider. Then run pnpm manasvi restart to apply changes to running services.
DeepSeek (default)
DeepSeek is the default model provider.
pnpm manasvi models add deepseek
Or manually in .env.local:
MANASVI_MODEL_PROVIDER=deepseek
MANASVI_MODEL=deepseek-v4-flash
MODEL_ADAPTER_MODE=deepseek
PLANNER_MODEL=deepseek-v4-flash
DEEPSEEK_API_KEY=your-key-here
DEEPSEEK_BASE_URL=https://api.deepseek.com
DEEPSEEK_TIMEOUT_MS=60000
Mock adapter
The mock adapter returns predictable test responses without connecting to any AI provider.
Useful for:
- Verifying the full pipeline works end-to-end before setting up a real model
- Automated tests
- Offline development
Ollama (local model)
Ollama runs models directly on your computer — no internet connection after setup, no API costs, data stays on your machine.
Install Ollama
Download from ollama.com and install it.
Download a model
# Llama 3.2 — good balance of speed and capability
ollama pull llama3.2
# Mistral — fast
ollama pull mistral
# Qwen — good instruction following
ollama pull qwen2.5
Configure via CLI
pnpm manasvi models add ollama
Or manually in .env.local:
MODEL_ADAPTER_MODE=ollama
OLLAMA_BASE_URL=http://localhost:11434/v1
PLANNER_MODEL=llama3.2
Start Ollama
ollama serve
Ollama needs to be running while Manasvi is running. Run pnpm manasvi doctor if connectivity fails — it checks whether the Ollama endpoint is reachable.
Smaller models (3B–7B parameters) run well on most modern laptops with 8GB RAM. Larger models (13B+) need more RAM or a GPU.
OpenAI (or compatible API)
pnpm manasvi models add openai
Or manually in .env.local:
MODEL_ADAPTER_MODE=openai
OPENAI_API_KEY=sk-your-key-here
PLANNER_MODEL=gpt-4o
OPENAI_BASE_URL=https://api.openai.com/v1
Which model? gpt-4o works well for most tasks. gpt-4o-mini is significantly cheaper and sufficient for most agent workflows.
OpenAI-compatible APIs: Change OPENAI_BASE_URL to point to any compatible service (Azure OpenAI, Groq, Mistral API, etc.).
Test connectivity
pnpm manasvi models test
This sends a minimal request and shows the error if the key is wrong or quota is exceeded.
Claude (Anthropic)
pnpm manasvi models add claude
Or manually in .env.local:
MODEL_ADAPTER_MODE=claude
ANTHROPIC_API_KEY=sk-ant-your-key-here
PLANNER_MODEL=claude-3-5-sonnet-latest
ANTHROPIC_BASE_URL=https://api.anthropic.com
During models add claude, Manasvi will:
- validate the API key (if possible),
- attempt model discovery through Anthropic Models API,
- let you choose a default Claude model,
- store the active provider cleanly alongside Ollama/OpenAI config.
You can obtain a key from console.anthropic.com.
Adjusting model behavior
These settings apply to all providers:
# How long to wait for the model (milliseconds)
MODEL_ADAPTER_TIMEOUT_MS=60000
# Maximum context chunks fed to the model
MODEL_ADAPTER_MAX_CONTEXT_CHUNKS=24
The timeout default is 60 seconds. Increase it if you're using a slow local model or a heavily loaded API.