agentwarden-ai 0.1.2

A local OpenAI-compatible proxy that reduces agent context waste and proves behavior held.

AgentWarden

AgentWarden is a local, OpenAI-compatible proxy for tool-using agents. It measures where each request spends input tokens, removes conservative dead weight from long-running conversations, and keeps a per-session savings receipt in SQLite.

Observability tools show you the bill. AgentWarden shrinks it and proves nothing broke.

What works today

• Transparent /v1/chat/completions proxy for OpenAI, including streaming.
• Per-request tracing: system prompt, tool schemas, conversation history, and current turn token counts.
• Independently enabled optimizers: unused tool pruning, old tool-output trimming, duplicate-history removal, and stable static-prefix ordering.
• Per-session budget warning, local SQLite traces, a demo coding agent, and a replay verifier that compares optimizations off versus on.

It supports applications that use the Chat Completions API. The proxy is language-neutral: a Python, Node.js, or other HTTP client can use it. The current v1 does not yet proxy OpenAI's /v1/responses endpoint or other model providers.

Quick start

Requirements: Python 3.11+ and an OpenAI API key. Your key stays in your application process and is forwarded directly to OpenAI; AgentWarden does not persist it.

Install from PyPI:

python3.11 -m venv .venv
.venv/bin/pip install agentwarden-ai

Until the first PyPI release is published, install the current GitHub version instead:

.venv/bin/pip install "git+https://github.com/Jahanshah1/AgentWarden.git"

Start the proxy with all current optimizers enabled:

export AGENTWARDEN_ENABLE_TOOL_PRUNE=true
export AGENTWARDEN_ENABLE_HISTORY_TRIM=true
export AGENTWARDEN_ENABLE_CONTEXT_DEDUP=true
export AGENTWARDEN_ENABLE_CACHE_ORDER=true
export AGENTWARDEN_SESSION_BUDGET_USD=0.02
.venv/bin/agentwarden serve

In a second terminal, check it is up:

.venv/bin/agentwarden doctor

The proxy listens on http://127.0.0.1:8080. Its OpenAI-compatible SDK base URL is http://127.0.0.1:8080/v1.

Add it to an existing agent

Keep the agent's OPENAI_API_KEY exactly where it already is. Run AgentWarden beside the agent, then point the OpenAI client at the local proxy.

Python:

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["OPENAI_API_KEY"],
    base_url="http://127.0.0.1:8080/v1",
)

Node.js:

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "http://127.0.0.1:8080/v1",
});

For useful per-agent receipts, send a stable session header with each agent run. This is optional; without it, traces use the default session.

client.chat.completions.create(
    model="gpt-5.6-terra",
    messages=messages,
    extra_headers={"X-AgentWarden-Session": "support-agent-run-42"},
)

Then inspect the local receipt:

curl "http://127.0.0.1:8080/stats?session_id=support-agent-run-42"
curl "http://127.0.0.1:8080/traces?session_id=support-agent-run-42"

Verify the project

For contributors working from a clone:

python3.11 -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/pytest -m 'not live' -q

To exercise a real OpenAI API key through the proxy:

export OPENAI_API_KEY="your-key"
.venv/bin/python scripts/smoke.py

To run the complete coding-agent demo and then inspect its receipt:

.venv/bin/agentwarden demo
.venv/bin/agentwarden stats --session-id demo-REPLACE_ME

For a separate consumer-agent proof that imports only the OpenAI SDK, run the independent support-agent example.

To compare the same demo task with optimizations off and on:

.venv/bin/agentwarden verify --no-judge

Configuration

All flags default to false, preserving byte-identical pass-through.

Variable	Meaning
AGENTWARDEN_ENABLE_TOOL_PRUNE	After three warm-up requests, sends only previously used or explicitly mentioned tools.
AGENTWARDEN_ENABLE_HISTORY_TRIM	Clips old tool-result messages while preserving recent turns.
AGENTWARDEN_ENABLE_CONTEXT_DEDUP	Replaces repeated old history content with a deterministic reference.
AGENTWARDEN_ENABLE_CACHE_ORDER	Makes the system-and-tools prefix stable to help provider prompt caching.
AGENTWARDEN_SESSION_BUDGET_USD	Adds a warning response header after the projected session cost crosses this amount.
AGENTWARDEN_DB_PATH	SQLite trace database location; defaults to agentwarden.sqlite3.

Dashboard

The dashboard is bundled with the Python package. No Node.js or separate frontend server is needed. Start it with:

agentwarden dashboard

Open http://127.0.0.1:8080/dashboard. The dashboard can select any recorded session, show before/after input context, inspect every request, and change optimizer flags or the budget-warning threshold for the currently-running proxy. Runtime changes take effect immediately but reset when the proxy is restarted; use environment variables for durable defaults.

Current status

This is a working hackathon prototype, not a hosted multi-tenant service. It is designed to run locally beside one developer's agent. Before a broad public release, the next work is the dashboard, packaged releases on PyPI, stronger live replay coverage, and support for the Responses API.