plap

An extensible OpenAI-compatible model harness

Responses and Chat Completions through one plugin-driven model loop.

Documentation · Write a plugin · Apache 2.0

plap moves model-harness behavior into the server:

• Python plugins can add tools and modify response execution through hooks.
• Threads let several models work with separate histories.
• Client tool calls and private plugin state can continue across requests.

Together, these capabilities let plap run multi-model workflows behind an OpenAI-compatible API without flattening every model's context into one transcript. Those workflows can survive client round trips too, as this review does:

Start plap

You need Pixi and Docker.

Create your local environment file:

cp .env.example .env

Add an OpenRouter key to .env:

OPENROUTER_API_KEY=your-key

Start the development server:

pixi run dev

This starts temporary PostgreSQL and telemetry containers, applies migrations, creates a development API key, and runs the server. The command prints the active model, URLs, and log path.

Keep it running. In another terminal, load the generated client settings:

source .dev/.env

Send a response

The OpenAI Python client is already installed in the Pixi environment. Start a Python session:

pixi run python

Point the client at plap:

import os

from openai import OpenAI

client = OpenAI(
    base_url=os.environ["PLAP_DEV_BASE_URL"],
    api_key=os.environ["PLAP_DEV_API_KEY"],
)

Responses

response = client.responses.create(
    model=os.environ["PLAP_DEV_MODEL"],
    input="Say hello in one sentence.",
)

print(response.output_text)

Chat Completions

completion = client.chat.completions.create(
    model=os.environ["PLAP_DEV_MODEL"],
    messages=[{"role": "user", "content": "Say hello in one sentence."}],
)

print(completion.choices[0].message.content)

You now have a working local plap server. To carry model and plugin state into later requests, continue with Continue a Conversation.

Add something

Write your first plugin to add a server_time tool that the model can call.

The documentation index covers the event bus, server tools, hooks, reasoning summaries, state, separate model contexts, and the lower-level LLM library.

Development commands

pixi run setup
pixi run pytest tests/unit
pixi run ruff check src tests scripts
pixi run ruff format --check src tests scripts

Tests marked money or expensive may call live providers and use credentials from root .env.

License

plap is licensed under the Apache License 2.0.