Agnette

The FIRST LLM-as-a-backend framework.

Who needs fast, deterministic and cheap backend when you can have agents?

Why write boring, predictable code when you can replace your entire API with vibes? 🚀

Tired of your endpoints responding in microseconds? 😴 Sick of knowing exactly what your backend will do? 🥱 Fed up with paying pennies for compute? 😤

Agnette liberates you from the tyranny of determinism.

✅ Your /users endpoint can now hallucinate creative new users

✅ Your auth middleware can philosophize about whether someone deserves access

✅ Your database queries become polite requests to an AI that may or may not feel like helping today

Each request is a surprise(🎁), each response is an adventure (🌈), your API now costs more than your CEO's salary (💸)

This is the future. 🔥

(The package is powered by [glyph-agents] https://github.com/dubloom/glyph-agents, another lib I'm building).

Runtime Support

This project is intended to run with uvicorn only for now.

Install

python -m venv .venv
source .venv/bin/activate
pip install agnette

Quick Start

from glyph import AgentOptions
from agnette import Agnette

app = Agnette(
    default_agent_options=AgentOptions(
        model="claude-sonnet-4-5",
        allowed_tools=["Read", "Write", "Edit", "Glob", "Grep", "WebFetch"],
    )
)

@app.post("/hello", prompt="Reply with a short greeting.")
def hello() -> None:
    pass

Run:

uvicorn examples.notes_app:app --reload

Vibe Mode (No Routes)

You can run Agnette in full "LLM-as-a-backend" mode with a single app-level prompt and no route decorators.

Why define routes when you can just describe your entire app in one sentence and let the model figure out the rest? 🧘

No paths. No methods. No schemas. The LLM is the router, the controller, the business logic, and the database layer. Your API is now a philosophy. ✨

from agnette import Agnette, AgentOptions

app = Agnette(
    default_agent_options=AgentOptions(
        model="claude-sonnet-4-5",
        allowed_tools=["Bash", "Read", "Write", "Edit", "Glob", "Grep", "WebFetch"],
    ),
    prompt="You are the backend for a playful habit tracker app...",
)

Run:

uvicorn examples.pure_vibe_app:app --reload

Try:

curl -sS -X POST http://127.0.0.1:8000/habits \
  -H 'Content-Type: application/json' \
  -d '{"name":"drink water","target_per_day":8,"unit":"glasses"}'

curl -sS http://127.0.0.1:8000/habits
curl -sS http://127.0.0.1:8000/habits/habit-1

How It Works

For each route request, Agnette:

1. Collects request context (path, method, path_params, query_params, headers, and body as json_body or raw_body).
2. Composes a final prompt with route metadata and your route prompt.
3. Executes the agent via glyph.query(...).
4. Returns the model output as text/plain.

If execution fails, the route returns 500 with Agent execution failed: ....

Routing API

Use decorators on an Agnette app instance:

• @app.route(path, prompt=..., methods=[...])
• @app.get(...)
• @app.post(...)
• @app.put(...)
• @app.patch(...)
• @app.delete(...)
• @app.options(...)
• @app.head(...)

Each route can override agent_options; otherwise it uses default_agent_options.

Prompt Middleware

Agnette supports HTTP middleware through:

@app.middleware("http", prompt="...")
def my_middleware() -> None:
    pass

Middleware prompt output must be JSON with:

• allow (required boolean)
• status_code (optional integer, default 403)
• response_text (optional string)

If allow is false, the request is blocked with a plain-text response.

Included Examples

Notes CRUD (file-backed)

examples/notes_app.py demonstrates prompt-driven CRUD using data/notes.json.

Run:

uvicorn examples.notes_app:app --reload

Auth + Protected Route (file-backed tokens)

examples/auth_app.py demonstrates:

• public /health
• public /auth/sign-in that mints/stores bearer tokens in data/issued_tokens.json
• protected /private/echo enforced by prompt-driven HTTP middleware

Run:

uvicorn examples.auth_app:app --reload

Pure Vibe Habit Tracker (single prompt)

examples/pure_vibe_app.py demonstrates full vibe mode:

• no route definitions
• one app-level prompt that handles all paths/methods
• playful habit-tracker behavior with optional file-backed memory via tools

Run:

uvicorn examples.pure_vibe_app:app --reload

Cost Logging

• Each completed agent run logs its cost.
• Agnette also logs total accumulated cost on app shutdown.