MCP server
Connect your agent to https://api.inferenceapis.com/mcp and it can manage an Inference APIs account by itself: create one for you, read prices, check the balance, see what a session cost, and confirm a key works. It cannot spend your money; adding funds stays on the billing page.
| Tool | Needs a key | What it does |
|---|---|---|
docs | No | The plain-text documentation (short or full) |
list_models | No | Models with prices, context length, capabilities and aliases; filter by type |
signup | No | Create an account and API key for a person. The $1 starter credit is released by the confirmation email |
account | Yes | Email, balance, confirmation and starter-credit status, where to add funds |
usage | Yes | Spend and tokens over the last N days, by model |
create_api_key | Yes | An additional key for a second tool or machine |
test_completion | Yes | One small chat request; returns the reply and what it cost |
Connect
The server speaks MCP over Streamable HTTP with JSON responses and keeps no session state. Send your API key as a Bearer token; tools that need no key work without one.
[mcp_servers.inferenceapis]
url = "https://api.inferenceapis.com/mcp"
bearer_token_env_var = "INFERENCE_API_KEY"claude mcp add --transport http inferenceapis https://api.inferenceapis.com/mcp --header "Authorization: Bearer $INFERENCE_API_KEY"{
"mcp": {
"inferenceapis": {
"type": "remote",
"url": "https://api.inferenceapis.com/mcp",
"headers": { "Authorization": "Bearer {env:INFERENCE_API_KEY}" }
}
}
}{
"mcpServers": {
"inferenceapis": {
"url": "https://api.inferenceapis.com/mcp",
"headers": { "Authorization": "Bearer ia-your-key" }
}
}
}curl -s https://api.inferenceapis.com/mcp -H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"list_models","arguments":{"type":"chat"}}}'No key yet?
Connect without a token and call signup with the person's email. It returns a key; put that key on the connection and the account tools start working. The same flow without MCP is POST /v1/signup, described in /llms.txt, and there is a setup skill for agents that would rather follow instructions than call tools.
Tested on 2026-09-18 with codex-cli 0.154.0 connected over HTTP: it called account, usage and test_completion and reported the balance, the day's spend and the test cost correctly. Spec version 2025-06-18; batch requests and server-initiated messages are not supported.