ra-yavuz › edgecall

edgecall

A small self-hosted, OpenAI-compatible API. You POST a sentence like "what is the gold price"; a deliberately weak local model (phi3-class, the kind that runs on a Raspberry Pi or a phone) reads a short menu of your registered functions, picks one, and edgecall runs it and returns the result.

The whole trick: the model never sees your full function list. Functions are shown a few at a time as a numbered menu, and the model pages through with a "next" option until it picks one. Picking one item off a short list is the easiest thing a weak model can do, so a weak model is enough. No embedding model, no index to maintain.

How it works

request: "i need to know the current gold price"
   |
   v
edgecall shows the weak model menu page 0:
   0001  show the message of the day
   0002  find the current gold price
   0003  get the current weather for a location
   NEXT  none of these, see more options
   |
   model replies: 0002
   v
edgecall runs function 0002 -> { "price": 4347.1, "currency": "USD", ... }

If the right function is not on the current page, the model replies NEXT and edgecall shows the next page. Pages are just slices of your function list in registration order. The trade-off: a function deep in the list costs the model one NEXT per page to reach, and weak models page blindly, so put your most-used functions first.

What it is for

You want useful pre-programmed actions triggered by natural language, but you do not want to pay for a frontier model or send anything to the cloud. edgecall lets a cheap local model do the routing while the actual work stays in plain Python functions you control. It does not host a model: it talks to any OpenAI-compatible /v1 endpoint you already run (llama.cpp, ollama, LM Studio, or a sibling hydra-llm server).

Add a function

Every *.py file in the functions directory is autoloaded at startup. Give it an id and a one-line description:

from edgecall.registry import register

@register(id="0006", desc="say hello to a name")
def run(args):
    name = args.get("name", "world")
    return {"greeting": f"hello, {name}"}

args["request"] is always the original sentence; other keys are whatever the caller passed. Return any JSON-serialisable value. The bundled examples include a message-of-the-day function served from a text file, plus live gold-price and weather lookups over public no-key APIs.

API

Install (Debian / Ubuntu)

sudo bash -c 'set -e; install -m 0755 -d /etc/apt/keyrings && curl -fsSL https://ra-yavuz.github.io/apt/pubkey.gpg -o /etc/apt/keyrings/ra-yavuz.gpg && echo "deb [signed-by=/etc/apt/keyrings/ra-yavuz.gpg] https://ra-yavuz.github.io/apt stable main" > /etc/apt/sources.list.d/ra-yavuz.list && apt update && apt install -y edgecall'

Then:

edgecall functions                            # list what is registered
edgecall serve                                # run the API on 127.0.0.1:8900
edgecall dispatch "what time is it" --trace   # route one request, no server

Run with Docker Compose

git clone https://github.com/ra-yavuz/edgecall
cd edgecall
# Edit docker-compose.yml: set EDGECALL_MODEL_BASE_URL to your endpoint.
docker compose up -d

curl -s -X POST localhost:8900/v1/dispatch \
     -H 'content-type: application/json' \
     -d '{"request":"what is the gold price"}'

The honest limitation

A weak model picking from a menu will sometimes pick wrong, and edgecall will faithfully run whatever it picked. That is the deal you accept for not paying for a frontier model. Mitigations baked in: a tight low-temperature reply, strict parsing (only an id actually on the page is accepted), bounded retries, and a full decision trace so you can see why a wrong call happened. For high-stakes actions, gate the function itself; do not rely on the model's pick alone.