How to Set Up OpenClaw with DeepSeek (Cheap Self-Hosted Agent)

OpenClaw gives you a personal assistant; DeepSeek gives it a cheap brain. That pairing is one of the most practical things you can do with OpenClaw, because the assistant’s whole appeal — running things on a schedule, looping on a task overnight, always being reachable — is also exactly how a premium model quietly drains your account. DeepSeek is one of the cheapest capable APIs around, so it turns an always-on assistant from a cost worry into a rounding error.

This guide connects OpenClaw to DeepSeek: the config, the key handling, which model to pick, and the gotchas. It assumes OpenClaw is already installed — if not, start with the Windows and WSL install guide.

Why this pairing works#

DeepSeek exposes an OpenAI-compatible API, which is the format OpenClaw already speaks. That means connecting it is configuration, not engineering — a base URL, a key, a model name. No proxy, no translation layer, no custom adapter. And because OpenClaw is model-agnostic by design, DeepSeek slots in as the brain without changing anything about how the assistant behaves.

Step 1: get a DeepSeek API key#

Create an account with DeepSeek and generate an API key from their platform. Note the base URL for the OpenAI-compatible endpoint and the model names from their integration docs. You’ll need three things: the base URL, the key, and a model name.

Step 2: set the key as an environment variable#

Keep the key out of config files. In your WSL or Linux shell:

export DEEPSEEK_API_KEY="sk-your-deepseek-key"

To make it persist across sessions, add that line to your shell profile (~/.bashrc or ~/.zshrc) and reopen the terminal.

Step 3: point OpenClaw at DeepSeek#

In OpenClaw’s configuration, set the model provider to DeepSeek’s OpenAI-compatible endpoint, reference the key from the environment, and pick a model. Config formats shift between OpenClaw versions, so confirm the current keys in the repo, but the values you’re supplying are always these:

A representative config block looks like this:

{
  "model": {
    "provider": "openai-compatible",
    "base_url": "https://api.deepseek.com",
    "api_key_env": "DEEPSEEK_API_KEY",
    "model": "deepseek-chat"
  }
}

The key idea is provider: openai-compatible plus DeepSeek’s base URL — that’s what tells OpenClaw to talk to DeepSeek in a format it already understands.

Step 4: pick the right model#

DeepSeek offers a general chat model and a reasoning model. The split is simple:

• General model — everyday assistant work: reading files, summaries, messaging, routine automation. Fast and cheapest.
• Reasoning model — harder, multi-step tasks where the assistant needs to think before acting. Slower and a bit pricier, but still cheap.

Set the general model as your default and switch to reasoning for the occasional hard job. If you later add routing, keep the cheap model as the default and send only hard tasks to the stronger model.

Step 5: test it#

Start OpenClaw and give it a task that uses both the model and a capability:

Summarize the README in this folder and tell me what the project does.

If it reads the file and answers sensibly, DeepSeek is wired in correctly. If the model replies but can’t read the file, the model connection is fine and the issue is a skill or permission — not DeepSeek.

A note on cost#

DeepSeek’s rates are low, but “low” still adds up if you put a chatty assistant on a tight schedule. Two habits keep it sane: use the general model as default and reserve the reasoning model for when it’s needed, and don’t schedule jobs more often than you actually use them. Prices on these models shift, so treat any figure you read as a snapshot and confirm on the official pricing page.

Wrapping up#

Connecting OpenClaw to DeepSeek is a config change, not a project: set DeepSeek’s OpenAI-compatible base URL, reference your key from an environment variable, pick a model, and test on a real task. The payoff is a self-hosted assistant you can leave running without watching the meter.

From here, if you’d rather pay a flat monthly rate, set up GLM instead. Want zero API cost? Go fully local with Ollama.