geniete-screens 0.1.0

GenieTé Screens Python SDK — live visual output for AI agents (AgentScreen)

geniete-screens

Publish live status, metrics, terminal output, and viewer links from any AI agent or automation into an AgentScreen.

AgentScreen is a GenieTé service by Genie, Inc.. It is built for developers who want their own systems to stay simple and predictable while offloading live connectivity, session handling, and fair metered delivery to a service made by developers.

Limited Access

AgentScreen currently runs as a limited-access service. You need:

• a GenieTé account at https://account.geniete.com/;
• a screen created at https://screens.geniete.com/;
• credits granted by an administrator;
• a screen API key, shown once when the screen is created.

If your balance reaches zero, ask your administrator for more credits.

Install

pip install geniete-screens

or:

uv add geniete-screens

The import name is geniete_screens.

Quick Start

Create a screen in the AgentScreen console, copy the API key, and note the screen id.

from geniete_screens import Screen

screen = Screen(
    api_key="as_key_...",
    screen_id="scr_...",
)

screen.message("Agent booted", title="status", tone="success")
screen.metric("queue_depth", 12)
screen.terminal("$ python agent.py\nready\n")

The default base URL is:

https://screens.geniete.com

Override it only for local or staging tests:

screen = Screen(
    api_key="as_key_...",
    screen_id="scr_...",
    base_url="https://staging.example.com",
)

Event Types

Message Card

screen.message(
    "Fetched 42 candidate documents and selected 5 for synthesis.",
    title="retrieval",
    tone="info",
)

Tones: info, success, warning, error.

Metric

screen.metric("latency", 184, unit="ms")
screen.metric("tokens", 18342)
screen.metric("cost", "0.42", unit="credits")

Calling again with the same metric name updates it in place.

Terminal Output

screen.terminal("running eval suite...\n")
screen.terminal("pass 128/128\n")

Clear The Screen

screen.clear()

Viewer Links

Create a viewer link when you want another person or system to watch a screen.

url = screen.create_viewer_link(ttl_seconds=3600)
print(url)

The server enforces the maximum TTL allowed by the current plan.

Active Sessions

for session in screen.sessions():
    print(session["session_id"], session.get("ip"))

screen.revoke_session("session_id_here")

Async Agents

Use AsyncScreen when your agent already runs in an async loop.

import asyncio
from geniete_screens import AsyncScreen

async def main():
    screen = AsyncScreen(api_key="as_key_...", screen_id="scr_...")
    await screen.message("Async agent started", tone="success")
    await screen.metric("steps", 1)
    sessions = await screen.sessions()
    print(f"{len(sessions)} active viewers")

asyncio.run(main())

Long-Running Streaming Agents

For high-frequency agent output, use the WebSocket producer stream.

import asyncio
from geniete_screens import AsyncScreen

async def run_agent():
    screen = AsyncScreen(api_key="as_key_...", screen_id="scr_...")

    async with screen.stream() as stream:
        await stream.message("planning")
        await stream.terminal("tool: web_search\n")
        await stream.metric("steps", 1)
        await stream.message("done", tone="success")

asyncio.run(run_agent())

Use With Agent Frameworks

AgentScreen does not require a specific agent framework. If your agent can run Python code or call an HTTP endpoint, it can publish to a screen.

Generic Agent Loop

from geniete_screens import Screen

screen = Screen(api_key="as_key_...", screen_id="scr_...")

def on_step(step_name: str, detail: str):
    screen.message(detail, title=step_name)

def on_tool_output(tool_name: str, output: str):
    screen.terminal(f"$ {tool_name}\n{output}\n")

def on_score(name: str, value: float):
    screen.metric(name, value)

LangChain, LlamaIndex, CrewAI, AutoGen, Or Custom Runners

Use the same pattern in the framework's callback, event handler, observer, or tool wrapper:

screen.message("tool call started", title="agent")
screen.terminal(tool_output)
screen.metric("retrieved_docs", len(documents))

Keep AgentScreen calls outside your critical decision path when possible. It is best used as live observability and customer-facing progress, not as part of the model's core reasoning state.

Shell, CI, Or Any HTTP Client

The SDK wraps this endpoint:

curl -X POST https://screens.geniete.com/public/api/publish \
  -H "Authorization: Bearer $AGENTSCREEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "screen_id": "scr_...",
    "event": {
      "type": "content.message",
      "title": "deploy",
      "body": "release finished",
      "tone": "success"
    }
  }'

Error Handling

httpx raises an exception for non-2xx responses. Common cases:

• 401: missing or invalid API key;
• 403: key does not allow publish for this screen;
• 404: screen not found;
• 413: event payload too large;
• 429: quota or credit exhaustion.

During limited access, a 429 usually means your account needs more credits from an administrator.

Publish Order

When releasing public packages:

The MCP package depends on the SDK.