Use Inference.net Models in Coding Agents

You do not need inf instrument, inf auth login, or Gateway headers. Use your Inference API key with base URL https://api.inference.net/v1. Model IDs come from GET /v1/models.

Set your API key

Create a key on the API Keys tab of the dashboard, then export it:

export INFERENCE_API_KEY=<your-api-key>

Restart your agent after you set the key. Some clients (OpenCode included) cache their provider configuration, and keep an empty key if the variable was missing on first load. The result is a Missing Authorization Bearer token error.

Get model settings from GET /v1/models

GET /v1/models needs no API key. It returns the values most clients ask for when you add a custom model:

curl -s https://api.inference.net/v1/models \
  | jq '.data[] | select(.id == "deepseek-v4.1-flash-1017")'

{
  "id": "deepseek-v4.1-flash-1017",
  "context_length": 262144,
  "max_completion_tokens": 131072,
  "reasoning_efforts": ["none", "low", "high", "max"],
  "input_modalities": ["text", "image"],
  "supported_endpoints": ["chat", "responses", "messages"]
}

Coding agents need tool calls. Most chat models support them (see Function Calling). llama-3.2-1b-instruct, llama-3.2-3b-instruct, and the inference-net/schematron-* extraction models do not. GET /v1/models does not report tool support.

OpenCode

Add a provider to opencode.json (in your project, or globally at ~/.config/opencode/opencode.json). OpenCode 2 uses the providers key. Older versions use provider.

{
  "$schema": "https://opencode.ai/config.json",
  "providers": {
    "inference": {
      "name": "Inference.net",
      "package": "@ai-sdk/openai-compatible",
      "settings": {
        "baseURL": "https://api.inference.net/v1",
        "apiKey": "{env:INFERENCE_API_KEY}"
      },
      "headers": {
        "x-inference-task-id": "opencode"
      },
      "models": {
        "deepseek-v4.1-flash-1017": {
          "name": "DeepSeek V4.1 Flash 1017",
          "limit": { "context": 262144, "output": 131072 },
          "capabilities": {
            "tools": true,
            "reasoning": {
              "supported": true,
              "toggle": true,
              "effort": ["low", "high", "max"]
            }
          }
        }
      }
    }
  }
}

limit comes from context_length and max_completion_tokens. effort comes from reasoning_efforts. toggle: true maps to none.

Claude Code

Claude Code uses the Anthropic Messages API. Use the base URL without /v1:

export ANTHROPIC_BASE_URL=https://api.inference.net
export ANTHROPIC_AUTH_TOKEN="$INFERENCE_API_KEY"
export ANTHROPIC_MODEL=deepseek-v4.1-flash-1017
export ANTHROPIC_DEFAULT_FABLE_MODEL=deepseek-v4.1-flash-1017
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4.1-flash-1017
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4.1-flash-1017
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4.1-flash-1017
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4.1-flash-1017
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=262144
claude

• Set every model variable. Claude Code sends background tasks, subagents, and /model choices to these slots. A slot you leave unset sends requests to a Claude model ID, which fails or bills at Claude prices.
• Set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the model's context_length. Otherwise Claude Code assumes 200K. For a model with a context of 1M or more, add [1m] to the model ID instead, for example moonshotai/kimi-k3[1m].
• Use ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. The API reads only Authorization: Bearer.

Codex

Add a provider to ~/.codex/config.toml:

model = "deepseek-v4.1-flash-1017"
model_provider = "inference"
model_context_window = 262144
model_max_output_tokens = 131072
model_reasoning_effort = "high"

[model_providers.inference]
name = "Inference.net"
base_url = "https://api.inference.net/v1"
env_key = "INFERENCE_API_KEY"
wire_api = "responses"
http_headers = { "x-inference-task-id" = "codex" }

Take model_context_window, model_max_output_tokens, and model_reasoning_effort from context_length, max_completion_tokens, and reasoning_efforts. For a model Codex does not know, it uses a fallback system prompt and tool set. fast codex on also writes a Codex model catalog so that Inference.net models get Codex's standard prompt and tools.

Other OpenAI-compatible clients

Cursor, Cline, Continue, and other clients with an "OpenAI-compatible" provider option need three values:

• Base URL: https://api.inference.net/v1
• API key: your Inference API key
• Model: an id from GET /v1/models, for a model that supports tool calls

Fill in context and output limits from the field table.

Optional headers

These headers work on every request. You do not need Gateway proxy headers (x-inference-provider, x-inference-provider-api-key, x-inference-provider-url) to call Inference.net models.

Troubleshooting

Missing Authorization Bearer token The client sent no key. Check that INFERENCE_API_KEY is set in the shell that starts the client, then restart the client. In Claude Code, this also happens when you set ANTHROPIC_API_KEY instead of ANTHROPIC_AUTH_TOKEN.

301 Moved Permanently, 404, or HTML instead of JSON The base URL is https://inference.net/v1. Change it to https://api.inference.net/v1.

The model is not found Copy the exact id from GET /v1/models. Some IDs include an owner prefix, such as moonshotai/kimi-k3.