Code (Agentic)

Agentic tool-use endpoint for coding assistants. Functions identically to /chat but additionally supports tool_use events in the SSE stream, enabling the client to execute tools on behalf of the model.

POST /code

This endpoint is the primary interface for the VSCode extension and other agentic coding clients. The model may request tool calls (file reads, searches, terminal commands, etc.), and the client is responsible for executing those tools and returning results in subsequent requests.

Note: This differs from the internal /internal/agent/code endpoint, where the server executes tools autonomously. On /code, tool execution is always delegated to the client.

---

Authentication

Requires one of:

• X-Api-Key header with a valid API key

• Authorization and X-Firebase-AppCheck headers for browser-based auth

---

Request Body

Parameter	Type	Required	Default	Description
model	string	No	Auto-routed	Model ID to use. Must be a model that supports tool use.
messages	array	Yes	--	Array of message objects, including tool results from prior turns.
stream	boolean	No	true	Whether to stream the response via SSE.
temperature	float	No	0.7	Sampling temperature. Range: 0.0 to 2.0.
max_tokens	integer	No	Model default	Maximum tokens to generate. Clamped to the model's maximum output limit.
session_id	string	No	--	Resume an existing conversation session.
tools	array	No	--	Tool definitions available for the model to call.

Message Object

Messages follow the Anthropic tool-use conversation format:

Field	Type	Required	Description
role	string	Yes	One of system, user, assistant, or tool.
content	string or array	Yes	Text content or array of content blocks (for tool results).
tool_use_id	string	Conditional	Required when role is tool. The ID of the tool call being responded to.

Tool Definition Object

Field	Type	Required	Description
name	string	Yes	Unique tool name (e.g., read_file, run_command).
description	string	Yes	Description of what the tool does.
input_schema	object	Yes	JSON Schema defining the tool's input parameters.

---

Request Example

Initial request with tool definitions:

curl -X POST https://inference-api-611798501438.us-central1.run.app/code \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: inf_your_api_key_here" \
  -d '{
    "model": "claude-sonnet",
    "messages": [
      {"role": "user", "content": "Read the contents of main.py and summarize it."}
    ],
    "tools": [
      {
        "name": "read_file",
        "description": "Read the contents of a file at the given path.",
        "input_schema": {
          "type": "object",
          "properties": {
            "path": {"type": "string", "description": "Absolute file path"}
          },
          "required": ["path"]
        }
      }
    ],
    "stream": true
  }'

Follow-up request with tool result:

curl -X POST https://inference-api-611798501438.us-central1.run.app/code \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: inf_your_api_key_here" \
  -d '{
    "model": "claude-sonnet",
    "session_id": "abc123",
    "messages": [
      {"role": "user", "content": "Read the contents of main.py and summarize it."},
      {"role": "assistant", "content": [
        {"type": "text", "text": "I will read the file for you."},
        {"type": "tool_use", "id": "tool_01", "name": "read_file", "input": {"path": "/app/main.py"}}
      ]},
      {"role": "tool", "tool_use_id": "tool_01", "content": "import fastapi\n\napp = FastAPI()\n\[email protected](\"/\")\ndef root():\n    return {\"status\": \"ok\"}"}
    ],
    "tools": [
      {
        "name": "read_file",
        "description": "Read the contents of a file at the given path.",
        "input_schema": {
          "type": "object",
          "properties": {
            "path": {"type": "string", "description": "Absolute file path"}
          },
          "required": ["path"]
        }
      }
    ],
    "stream": true
  }'

---

Response (Streaming)

The SSE stream may include the following event types:

chunk -- Incremental text content

data: {"type": "chunk", "content": "I'll read the file for you."}

tool_use -- Model requests a tool call

data: {"type": "tool_use", "id": "tool_01", "name": "read_file", "input": {"path": "/app/main.py"}}

When the client receives a tool_use event, it should:

1. Execute the requested tool locally.

2. Send a new request to /code with the tool result appended to the messages array.

3. Continue until the model produces a final text response without further tool calls.

done -- Stream complete

data: {"type": "done", "model": "claude-sonnet", "session_id": "abc123", "credits_used": 34, "input_tokens": 128, "output_tokens": 210}

error -- Error during generation

data: {"type": "error", "message": "Model returned an error"}

---

Agentic Conversation Loop

A typical agentic session follows this pattern:

Client                              Server
  |                                    |
  |  POST /code (user message)         |
  |  --------------------------------> |
  |                                    |
  |  SSE: chunk, chunk, tool_use       |
  |  <-------------------------------- |
  |                                    |
  |  [Client executes tool locally]    |
  |                                    |
  |  POST /code (with tool result)     |
  |  --------------------------------> |
  |                                    |
  |  SSE: chunk, chunk, done           |
  |  <-------------------------------- |

Each round-trip within the same session is billed separately. The session_id returned in the done event should be passed in subsequent requests to maintain conversation context.

---

Credit Billing

Credit billing follows the same reservation-and-reconciliation model as /chat. Each round-trip (request/response pair) is billed independently based on the tokens consumed in that exchange.

---

Error Responses

Status	Description
400	Invalid request body, parameters, or tool definitions
401	Missing or invalid authentication
403	Insufficient credits or selected model does not support tool use
404	Specified model not found
429	Rate limit exceeded
500	Provider or internal server error