API and opencode

The gateway speaks the OpenAI chat-completions API at https://api.aisathome.org/v1. Any client that lets you set a base URL and a key works: opencode, the OpenAI SDKs, curl.

Get a key

Sign in with GitHub or Google on your account page and create a key there (up to three). A key is shown once; copy it then. Keys are personal and share your account’s daily limit with the chat. If a key leaks, revoke it on the same page.

Models

Name	What
deep	Qwen3.8 27B, general chat
agent	The same model, for coding agents and tool calls

GET /v1/models lists what the network serves right now.

Try it with curl

export AISATHOME_API_KEY=ak-...
curl https://api.aisathome.org/v1/chat/completions \
  -H "Authorization: Bearer $AISATHOME_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"deep","messages":[{"role":"user","content":"What is BOINC?"}]}'

opencode

Add a provider to ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "aisathome": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "AIs@home",
      "options": {
        "baseURL": "https://api.aisathome.org/v1",
        "apiKey": "{env:AISATHOME_API_KEY}"
      },
      "models": {
        "agent": {
          "name": "AIs@home agent (Qwen3.8 27B)",
          "limit": { "context": 32768, "output": 8192 }
        }
      }
    }
  }
}

Set context to the context_length that /v1/models reports for agent (the context every node serving it can take, for example 32768):

curl -s https://api.aisathome.org/v1/models -H "Authorization: Bearer $AISATHOME_API_KEY"

Then export AISATHOME_API_KEY=ak-... and run opencode -m aisathome/agent. Tool calls (reading and editing files, running commands) work.

Good to know

• Secrets: a safety net, not a guarantee. A simple scan tries to replace well-known key formats with a placeholder like sk-proj-AISX…XXXX before a node sees them, and does not put them back into the answer or into tool calls, so a file the agent writes may contain the placeholder instead of your real key. Unusual formats get through: treat the service as untrusted and keep secrets out of what the agent can read.
• Who serves you. Chats and plain API requests may be answered by any volunteer’s node; anything that looks like personal data goes only to the maintainer’s nodes. Requests with tools (agents) go to volunteers’ nodes when the key allows it: Run requests through nodes operated by other people is ticked by default when you create a key. If you untick it, agent requests go only to your own node or the maintainer’s, which are often offline.
• Run agents carefully, with any provider. A model can propose a harmful command, and a malicious node could alter one; volunteers’ answers pass simple checks but are not verified yet. In opencode, keep approval of commands on, and run it isolated: in a dedicated Docker container or virtual machine with only the project mounted, no SSH keys, cloud credentials or browser profile, and nothing you could not afford to lose.
• Context and answer length. A request must fit the context of a node that serves it: the prompt (everything sent, including tool definitions and earlier turns) plus the room for the answer. /v1/models reports context_length, the context every node serving the model can take right now. A request larger than any node can take fails at once with 400 context_length_exceeded, naming the largest available. Answers are capped at 32,768 tokens; without max_tokens, at 16,384.
• Limits: a daily token budget, requests per minute and concurrent requests per key. Over the limit you get 429 with a Retry-After header.