Agent setup examples
Create a chat agent key and configure OpenAI-compatible agents, Hermes, the OpenAI SDK, or cURL.
Updated
Use a llmbase_chat_... key with the agent API base URL when an external agent
should consume your Pro Chat/Agent subscription budget.
Create a chat agent key
- Open Dashboard -> Agent Access.
- In Chat agent keys, create a new key.
- Copy the key immediately. It is only shown once.
Revoking a key disables it immediately. If a bounded-lifetime key is shown as expired, create a replacement instead of continuing to use the old token.
Configure an OpenAI-compatible agent
Use these settings in any agent that supports a custom OpenAI-compatible API:
| Setting | Value |
|---|---|
| Base URL | https://llmbase.ai/api/v1/agents |
| API key | Your llmbase_chat_... chat agent key |
| Models endpoint | GET /models |
| Chat endpoint | POST /chat/completions |
Chat completions accept a JSON body of up to 8 MiB (8,388,608 bytes). A larger
body returns HTTP 413 with code request_body_too_large and error.type
request_too_large. See
Usage and rate limits for how that cap differs from
the included budget and from rate limits.
Example OpenCode configuration
Use a custom OpenAI-compatible provider. Set limit.context to the model’s
context window so OpenCode compacts from token usage before a request reaches
the 8 MiB (8,388,608-byte) body limit. DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash) has a
1,048,576-token context window, so set limit.context to 1048576. Set
limit.output to 384000, the model’s maximum output length.
compaction.auto turns on automatic compaction, including one compact-and-retry
when this API returns HTTP 413 with error.type request_too_large.
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"llmbase": {
"npm": "@ai-sdk/openai-compatible",
"name": "LLMBase",
"options": {
"baseURL": "https://llmbase.ai/api/v1/agents",
"apiKey": "{env:LLMBASE_CHAT_AGENT_KEY}"
},
"models": {
"deepseek/deepseek-v4.1-flash": {
"name": "DeepSeek V4.1 Flash",
"limit": {
"context": 1048576,
"output": 384000
}
}
}
}
},
"compaction": {
"auto": true
}
}
Run /connect, choose Other, and use the provider id llmbase. The key can
also live in LLMBASE_CHAT_AGENT_KEY, as in the snippet.
OpenCode V1 only: also set compaction.prune to true. That removes old tool
outputs. V2 summarizes older conversation and does not use prune.
{
"compaction": {
"auto": true,
"prune": true
}
}
Example Hermes configuration
Hermes should be configured as a named custom provider. A bare custom
provider with only OPENAI_API_KEY may not send the key to non-OpenAI hosts.
Load the current agent model list first, then choose a returned model with the
tool, context, and cost profile your agent loop needs.
model:
provider: llmbase
model: <model-id-from-agent-models>
base_url: https://llmbase.ai/api/v1/agents
api_key: llmbase_chat_...
api_mode: chat_completions
custom_providers:
- name: llmbase
base_url: https://llmbase.ai/api/v1/agents
api_key: llmbase_chat_...
api_mode: chat_completions
model: <model-id-from-agent-models>
Hermes uses reasoning_effort: medium by default. LLMBase accepts that default
for reasoning-capable models. When a model publishes high but no native
medium tier, the request uses high automatically; no Hermes override is
required. Explicit unsupported values still return a validation error.
Example with the OpenAI SDK
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://llmbase.ai/api/v1/agents",
apiKey: process.env.LLMBASE_CHAT_AGENT_KEY,
});
const models = await client.models.list();
const model = models.data[0]?.id;
const response = await client.chat.completions.create({
model,
messages: [
{ role: "user", content: "Draft a short product update." },
],
});
console.log(response.choices[0]?.message?.content);
Example cURL request
curl https://llmbase.ai/api/v1/agents/chat/completions \
-H "Authorization: Bearer $LLMBASE_CHAT_AGENT_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "<model-id-from-agent-models>",
"messages": [
{ "role": "user", "content": "Summarize this repo issue." }
]
}'