Cephable Automate HTTP Server

Embed a complete on-device agent.
It brings its own tools. It runs yours.

Teams building AI agents can now drop Cephable's on-device assistant into their own apps — browser research, computer use, document generation, long-form content, human-in-the-loop — behind a single local HTTP call. And it calls your functions back, in your process, against your data.

  • 46 built-in tools
  • Runs your functions
  • Nothing leaves the machine
Your app or agent
python · langchain
Contoso Deal Desk
YOUR APP · YOUR TOOLS
@tool
def get_internal_pricing():
    return warehouse.query(PRICING)

cephable.run(prompt, tools=[get_internal_pricing])
WHAT YOUR APP SEES
The local Cephable agent
on this device · no network
Cephable AI Workflows
Using tools: get_internal_pricing +46 built in
Research our top 3 competitors on pricing, then write it up as a deck.
Making a plan…
Ask Cephable to do more with this result…
Automate ▾
No new SDK to learn

Point the stack you already use at it

The Automate server speaks OpenAI's chat-completions API. For most code you have already written, becoming an on-device agent is a base URL and a key.

Before · someone else's GPU
client = OpenAI(
    base_url="https://api.openai.com/v1",
    api_key=os.environ["OPENAI_API_KEY"],
)
After · the agent on this device
client = OpenAI(
    base_url="http://127.0.0.1:4317/v1",
    api_key=os.environ["CEPHABLE_AUTOMATE_KEY"],
)
  • Already works with
  • OpenAI SDK
  • LangChain
  • Vercel AI SDK
  • MCP servers
  • .NET
  • curl, or any HTTP client
Already in the box

You are not building the hard part again

Every one of these ships today, runs on the device, and is reachable from the same single call.

Browser research Drives a real browser in the background and reads pages back as structured content. Every source stays addressable.
Computer use Enumerates the UI through the OS accessibility tree, clicks by name, reads the open document, types, sends shortcuts, launches apps.
Real documents Word, PowerPoint, Excel and PDF in and out — and edited in place, not regenerated from scratch.
Long-form content Generation, summarizing and translation chunk and recombine, so document length isn't capped by the context window.
Your tools, your process Declare a function; the agent calls it and waits while you run it against your own data and credentials. Nothing of yours has to be reachable.
Asks, and asks permission Pauses to put a real question to the user. Deletes, moves and shell commands stop for approve, edit or reject.
46built-in tools, before you add yours
1local HTTP call to start a run
0bytes off this device
Open source

Samples you can read and run

Each one runs against a fake Cephable included in the repo, so you can read the code and watch the protocol work before installing anything. More are on the way.

Python LangChain with your tools A support-triage agent running against your own data. Sixty lines of real integration. WinUI Windows AI, side by side Windows OCR reads it, Phi Silica rewrites it, Cephable decides. Three AI systems, no network. Next.js Vercel AI Elements chat A production-shaped chat UI with tool-call cards, pointed at the local assistant. Node Public gateway Auth, rate limits and an allow-list in front of the local port. Interesting for what it refuses.
Browse the samples Read the docs