Subconscious SDK

The official Python SDK for the Subconscious API

Installation

uv add subconscious-sdk
# or
pip install subconscious-sdk

Note: The package name is subconscious-sdk but you import it as subconscious.

Quick Start

from subconscious import Subconscious

client = Subconscious(api_key="your-api-key")

run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Search for the latest AI news and summarize the top 3 stories",
        "tools": [{"type": "platform", "id": "fast_search"}],
    },
    options={"await_completion": True},
)

print(run.result.answer)

Get Your API Key

Create an API key in the Subconscious dashboard.

Usage

Run and Wait

The simplest way to use the SDK—create a run and wait for completion:

run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Analyze the latest trends in renewable energy",
        "tools": [{"type": "platform", "id": "fast_search"}],
    },
    options={"await_completion": True},
)

print(run.result.answer)
print(run.result.reasoning)  # Structured reasoning nodes

Fire and Forget

Start a run without waiting, then check status later:

run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Generate a comprehensive report",
        "tools": [],
    },
)

print(f"Run started: {run.run_id}")

# Check status later
status = client.get(run.run_id)
print(status.status)  # 'queued' | 'running' | 'succeeded' | 'failed' | 'canceled' | 'timed_out'

Poll with Custom Options

run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Complex task",
        "tools": [{"type": "platform", "id": "fast_search"}],
    },
)

# Wait with custom polling options
result = client.wait(
    run.run_id,
    options={
        "interval_ms": 2000,  # Poll every 2 seconds
        "max_attempts": 60,   # Give up after 60 attempts
    },
)

Streaming (Text Deltas)

Stream text as it's generated:

for event in client.stream(
    engine="tim-claude",
    input={
        "instructions": "Write a short essay about space exploration",
        "tools": [{"type": "platform", "id": "fast_search"}],
    },
):
    if event.type == "delta":
        print(event.content, end="", flush=True)
    elif event.type == "done":
        print(f"\n\nRun completed: {event.run_id}")
    elif event.type == "error":
        print(f"Error: {event.message}")

Note: Rich streaming events (reasoning steps, tool calls) are coming soon. Currently, the stream provides text deltas as they're generated.

Skills

Attach reusable knowledge packages to your runs. Skills use progressive disclosure: the agent sees a summary in its system prompt and loads full instructions on demand.

run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Build a REST API following our team standards",
        "tools": [{"type": "platform", "id": "web_search"}],
        "skills": ["api-design", "error-handling"],
    },
    options={"await_completion": True},
)

Skills are resolved by name. You can browse and create skills at subconscious.dev/platform/skills or via the Skills API.

Structured Output

Get responses in a specific JSON schema format using Pydantic models:

from pydantic import BaseModel
from subconscious import Subconscious

class AnalysisResult(BaseModel):
    summary: str
    key_points: list[str]
    sentiment: str

client = Subconscious(api_key="your-api-key")

run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Analyze the latest news about electric vehicles",
        "tools": [{"type": "platform", "id": "fast_search"}],
        "answerFormat": AnalysisResult,  # Pass the Pydantic class directly
    },
    options={"await_completion": True},
)

# The answer will conform to your schema
print(run.result.answer)  # JSON string matching AnalysisResult

The SDK automatically converts your Pydantic model to JSON Schema. You can also pass a raw JSON Schema dict if preferred.

Multimodal Input (Images)

Send images alongside text instructions using the Image helper:

from subconscious import Subconscious, Image

client = Subconscious(api_key="your-api-key")

# From a local file
run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Describe what you see in this image",
        "content": [Image.from_path("photo.png")],
    },
    options={"await_completion": True},
)

print(run.result.answer)

Image supports several sources:

# From a URL (server-side fetch)
Image.from_url("https://example.com/photo.jpg")

# From a URL (client-side fetch, embeds as base64)
Image.from_url("https://example.com/photo.jpg", fetch=True)

# From raw bytes
Image.from_bytes(image_bytes)

# From a previously uploaded blob
Image.from_blob_ref(blob_key="abc123", mime="image/png")

Supported formats: PNG, JPEG, GIF, and WebP.

Image Tool Responses

If you have a function tool that generates or fetches an image, return it as a ToolResponse with mixed text and image content. The agent will see the image and can reason about it. This is useful for tools that generate images, such as a screenshot tool for computer use.

from subconscious import ToolResponse, Image

# Your tool endpoint receives a tool_call_id from the agent.
# Return an image alongside text using ToolResponse.build():
response = ToolResponse.build(
    tool_call_id="call_abc123",
    content=[
        "Here is the generated chart:",
        Image.from_path("chart.png"),
    ],
)

# Or from a URL
response = ToolResponse.build(
    tool_call_id="call_abc123",
    content=[
        "Screenshot captured successfully.",
        Image.from_url("https://example.com/screenshot.png"),
    ],
)

ToolResponse.build() accepts a string, a single content block, or a mixed list of strings and Image blocks — it normalizes everything into the wire format automatically. You can use this when building custom tools to use with Subconscious.

Tools

Platform tools — Hosted search and retrieval tools like fast_search, web_search, news_search, and more. See the tools documentation for the full list.

# Platform tools (hosted by Subconscious)
fast_search = {
    "type": "platform",
    "id": "fast_search",
}

# Function tools (your own HTTP endpoints)
custom_function = {
    "type": "function",
    "name": "get_weather",
    "description": "Get current weather for a location",
    "url": "https://api.example.com/weather",
    "method": "GET",
    "timeout": 30,
    "parameters": {
        "type": "object",
        "properties": {
            "location": {"type": "string"},
        },
        "required": ["location"],
    },
}

# MCP tools (connect to any MCP server)
mcp_tool = {
    "type": "mcp",
    "url": "https://mcp.example.com",
    "allowedTools": ["search", "get_page"],
}

Tool Headers & Default Arguments

Function tools support two powerful features for injecting data at call time:

• **headers**: HTTP headers sent with the request to your tool endpoint
• **defaults**: Parameter values hidden from the model and injected automatically

tool_with_headers_and_defaults = {
    "type": "function",
    "name": "search_database",
    "description": "Search the database",
    "url": "https://api.example.com/search",
    "method": "POST",
    "parameters": {
        "type": "object",
        "properties": {
            "query": {"type": "string", "description": "Search query"},
            # Define these for validation, but they'll be hidden from the model
            "session_id": {"type": "string"},
            "api_key": {"type": "string"},
        },
        "required": ["query"],  # Only query is required - model generates this
    },

    # HEADERS: Sent as HTTP headers when this tool's endpoint is called
    "headers": {
        "x-custom-auth": "my-secret-token",
        "x-request-source": "my-app",
    },

    # DEFAULTS: Injected into parameters, hidden from model
    "defaults": {
        "session_id": "user-session-abc123",
        "api_key": "secret-api-key",
    },
}

How it works:

Default arguments flow:

1. Define all parameters in properties (required for validation)
2. Parameters with defaults are stripped from the schema before the model sees them
3. Model only generates values for non-defaulted parameters (e.g., query)
4. At call time, defaults are merged into the request body
5. Default values always take precedence over model-generated values

Each tool can have its own headers and defaults - they're only applied when that specific tool is called.

MCP Tools

Connect to any Model Context Protocol server and use its tools in your runs. Subconscious discovers tools from the server, filters by your allowedTools list, and proxies calls automatically.

Authentication

MCP servers that require authentication accept an auth object. The auth translates to an HTTP header sent with every tool call:

from subconscious import Subconscious, MCPTool, McpAuth

client = Subconscious()

# Basic — use all tools from an MCP server
run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Find my recent meeting notes",
        "tools": [
            MCPTool(url="https://mcp.notion.so/v1"),
        ],
    },
    options={"await_completion": True},
)

# Filter to specific tools
run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Search my documents",
        "tools": [
            MCPTool(
                url="https://mcp.notion.so/v1",
                allowed_tools=["search", "get_page"],  # case-insensitive
            ),
        ],
    },
    options={"await_completion": True},
)

# With bearer auth (most common — e.g. OAuth tokens)
# → sends header: { "Authorization": "Bearer <token>" }
run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Check my calendar",
        "tools": [
            MCPTool(
                url="https://mcp.google.com/v1",
                auth=McpAuth(type="bearer", token="your-oauth-token"),
            ),
        ],
    },
    options={"await_completion": True},
)

# API key auth with custom header
# → sends header: { "X-Api-Key": "<token>" }
# The header name is typically "X-Api-Key" but may vary —
# check the docs of the MCP server you are connecting to.
run = client.run(
    engine="tim-claude",
    input={
        "instructions": "Query the database",
        "tools": [
            MCPTool(
                url="https://mcp.example.com",
                auth=McpAuth(type="api_key", token="key123", header="X-Api-Key"),
            ),
        ],
    },
    options={"await_completion": True},
)

**allowedTools filtering:**

You can also pass MCP tools as plain dicts:

{"type": "mcp", "url": "https://mcp.example.com", "allowedTools": ["search"]}

Error Handling

from subconscious import (
    Subconscious,
    SubconsciousError,
    AuthenticationError,
    RateLimitError,
)

try:
    run = client.run(...)
except AuthenticationError:
    print("Invalid API key")
except RateLimitError:
    print("Rate limited, retry later")
except SubconsciousError as e:
    print(f"API error: {e.code} - {e}")

Webhooks

Get a POST when runs complete instead of polling.

Per-run callback: pass callbackUrl on any run:

run = client.run(
    engine="tim-claude",
    input={"instructions": "Generate a report"},
    output={"callbackUrl": "https://your-server.com/webhook"},
)

Org-wide subscriptions: receive webhooks for all runs. Manage in the dashboard or via the API:

curl -X POST https://api.subconscious.dev/v1/webhooks/subscriptions \
  -H "Authorization: Bearer $SUBCONSCIOUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "callbackUrl": "https://your-server.com/webhook",
    "eventTypes": ["job.succeeded", "job.failed"],
    "secret": "your-signing-secret"
  }'

Subscriptions support enable/disable, HMAC-SHA256 signing, and a delivery log. See the webhooks docs for more.

Cancellation

# Cancel a running run
client.cancel(run.run_id)

API Reference

See the full API documentation for detailed reference. The Subconscious client exposes five core methods: run(), stream(), get(), wait(), and cancel().

For available engines and pricing, see the pricing page.

A run's status field is one of: queued, running, succeeded, failed, canceled, or timed_out.

Development

We recommend uv as your package manager. See pyproject.toml for Python version requirements and dependencies.

Upgrading

If you're upgrading from 0.x, see the Migration Guide for breaking changes and code examples.

Contributing

Contributions are welcome! Please feel free to submit a pull request.

Support

For support and questions:

• Documentation: https://docs.subconscious.dev
• Email: {hongyin,jack,dana,wei}@subconscious.dev