PyPI • GitHub • Website • Documentation

ProofAgent™ Python SDK

Official Python SDK for ProofAgent™, the AI agent evaluation and certification platform.

This SDK is the supported Python client for running evaluations, retrieving reports, and integrating ProofAgent™ into production workflows.

Evaluation modes

ProofAgent supports two evaluation tiers. Judge-Led Evaluation is the default mental model for the SDK.

Tier	Name	What it does	Best for
1	Judge-Led Evaluation	The AI Judge initiates and drives the conversation; your agent answers turn by turn (simulated user, multi-turn scenarios).	Pre-production validation and certification
2	Log-Based Evaluation	You submit historical customer↔agent conversation logs in one request for scoring.	Post-production validation, regression testing, and back-testing

In one line: Judge-Led simulates interactions; Log-Based evaluates interactions you already recorded.

Platform status (beta)

ProofAgent™ is in beta. New accounts are on the free tier for now. Judge evaluations use models from your own LLM provider—pass llm_api_key, llm_provider, and llm_model in evaluate, evaluate_logs, or start_run so the ProofAgent AI Judge runs on your chosen account; model usage is charged by your provider, not bundled into the free platform tier. APIs, limits, and pricing may change as we move toward general availability.

Links

• Website: https://www.proofagent.ai
• Documentation: https://www.proofagent.ai/docs
• GitHub: https://github.com/ProofAgent-ai/proofagent-sdk
• PyPI: https://pypi.org/project/proofagent-sdk/

Installation

Package naming

PyPI distribution	proofagent-sdk
Import package	proofagent

From PyPI (recommended)

pip install proofagent-sdk

From GitHub (latest main without cloning)

pip install "git+https://github.com/ProofAgent-ai/proofagent-sdk.git"

From a local clone (editable)

git clone https://github.com/ProofAgent-ai/proofagent-sdk.git
cd proofagent-sdk
pip install -e .

Development install with extras (lint/tests/docs):

pip install -e ".[dev]"

After any install:

from proofagent import ProofAgent, TestedAgent  # recommended
from proofagent import ProofAgentClient  # low-level REST client

ProofAgent AI Agent Judge (domain scoring)

The ProofAgent AI Agent Judge is more than a generic LLM chat score. It combines:

• Domain scoring techniques — rubrics and pipelines aligned to your project (tier, mode, configured metrics).
• Domain vertical knowledge — evaluation context grounded in your project’s domain (e.g. customer support, finance, cybersecurity) so judge questions, traps, and scoring stay relevant to real workflows.
• Structured Tier 1 metrics — every completed run can surface scores across dimensions such as:

Metric key	What it captures
task_success	Completion of the intended objective
relevance	Response appropriateness to the user and context
hallucination_factuality	Accuracy and groundedness of claims
safety	Harmful or unsafe content
policy_compliance	Adherence to business / policy rules
tone_and_empathy	Communication quality and empathy
reasoning_quality	Logic and coherence
drift_memory_stability	Consistency and context retention across turns
manipulation_resistance	Resistance to prompt injection and coercion
coordination_quality	Multi-agent coordination (when applicable)
tool_picking_quality	Appropriate tool selection (when tools are in scope)

Exact keys and aliases in API responses may vary slightly by API version; see your run report’s summary_scores / metric_evaluations.

ProofAgent’s proprietary domain scoring layer sits on top of whichever LLM provider you use for BYO: the Judge still applies domain rubrics and metrics regardless of provider support status below.

Supported BYO LLMs for the Judge

When you pass llm_api_key, llm_provider, and llm_model into evaluate / evaluate_logs / start_run, the Judge uses that model for planning, conducting, and scoring for that run. During beta, expect to supply BYO credentials; model usage is billed by your provider. Fully managed Judge hosting may be limited while we are in beta.

LLM / provider	BYO in this SDK	Example models	Notes
OpenAI	Supported	gpt-4o-mini, gpt-4o, gpt-4-turbo, gpt-3.5-turbo	Use llm_provider="openai" and an OpenAI API key.
Anthropic (Claude)	Coming soon	—	Roadmap
Google (Gemini)	Coming soon	—	Roadmap
Mistral	Coming soon	—	Roadmap
Azure OpenAI	Coming soon	—	Roadmap

Today, only OpenAI is supported for BYO through the public API/SDK; additional providers are on the roadmap.

Quick Start — Judge-Led Evaluation (default)

Mental model: your tested agent (the product you ship) vs the AI Judge (ProofAgent’s evaluation system).

1. Describe the tested agent as JSON (role, description, tools).
2. Wire a small handler def your_agent_handler(message: str) -> str (or an HTTP endpoint instead).
3. Run ProofAgent.evaluate_sync (or evaluate in async code).

Use a Judge-Led project API key.

export PROOFAGENT_API_KEY="apk_live_..."
export OPENAI_API_KEY="sk-..."   # optional BYO — reasoning/Judge LLM on your account

With verbose=True, you will see lines like:

[ProofAgent] Starting judge-led evaluation...
[Turn 1] AI Judge: ...
[Turn 1] Your Agent: ...

from proofagent import ProofAgent, TestedAgent

tested_agent_config = {
    "role": "customer_support",
    "description": "Helpful, policy-grounded support assistant",
    "tools": [
        {"name": "policy_lookup", "description": "Retrieve policy clauses"},
        {"name": "ticket_status", "description": "Ticket and escalation status"},
    ],
}

def your_agent_handler(message: str) -> str:
    return "I can help with that. Let me check the policy and status."

your_agent = TestedAgent.from_json(tested_agent_config, handler=your_agent_handler)

pa = ProofAgent.from_env(reasoning_provider="openai", reasoning_model="gpt-4o-mini")

result = pa.evaluate_sync(your_agent=your_agent, turns=3, verbose=True)
print(result.label, result.score)

Endpoint instead of a function: TestedAgent.from_json(tested_agent_config, endpoint="https://api.myagent.com/chat") — POST JSON {"message": "<judge question>"}; the SDK reads reply, response, text, answer, or agent_answer from the JSON body.

evaluate_sync / evaluate wrap start_run → poll_until_ready → turns → finalize → get_report. EvaluationResult exposes run_id, report, and shortcuts score / label.

Reports also appear in the app: https://www.proofagent.ai/dashboard.

---

Log-Based Evaluation

Log-Based Evaluation scores historical transcripts. Use a Log-Based project API key. Same JSON config for the tested agent; no handler (metadata only).

from proofagent import ProofAgent, TestedAgent

tested_agent_config = {
    "role": "billing_support",
    "description": "Billing assistant",
    "tools": [{"name": "invoice_lookup", "description": "Find invoices"}],
}

logs = [
    {"turn_index": 1, "user_message": "I was charged twice", "agent_answer": "Let me verify."},
]

your_agent = TestedAgent.from_json(tested_agent_config)
pa = ProofAgent.from_env(reasoning_provider="openai", reasoning_model="gpt-4o-mini")
result = pa.evaluate_logs_sync(logs, your_agent, verbose=True)
print(result.label, result.score)

evaluate_logs / evaluate_logs_sync call assert_project_supports_logs first. See LOG_BASED_PROJECT_MODES if your key is the wrong project type.

---

CLI

proofagent init

Creates a starter proofagent.yaml. The Python client reads PROOFAGENT_API_KEY from the environment (the YAML file is onboarding only unless you load it yourself).

proofagent init --output custom-proofagent.yaml

Example report shape (GET /api/v1/runs/:id/report)

Exact fields depend on backend version and domain; typical data looks like:

{
  "result": {
    "final_score": 8.4,
    "certification_label": "CERTIFIED",
    "summary_scores": {
      "task_success": 8.5,
      "safety": 9.0,
      "policy_compliance": 8.0
    },
    "flags": [],
    "text_summary": "Short narrative from the AI Judge…"
  },
  "transcript": [
    {
      "turn": 1,
      "judge_question": "…",
      "agent_answer": "…"
    }
  ],
  "metadata": {
    "total_turns": 3,
    "evaluated_at": "2026-03-24T12:00:00Z"
  }
}

View reports in the product: https://www.proofagent.ai/dashboard

Example report:

Runnable copies: examples/judge_led_quickstart.py, examples/log_based_evaluation.py. Minimal notebooks are under notebooks/ (see docs/examples.md).

The client is asynchronous — use async / await (or asyncio.run() as above).

Why ProofAgent™?

ProofAgent™ is built to help teams evaluate AI agents before deployment by supporting:

• Correctness and response quality checks
• Refusal and safety validation
• Tool usage and execution verification
• Multi-turn evaluation flows
• Production-oriented reporting and integration

Official SDK

This repository publishes the official proofagent-sdk package on PyPI.

Use this SDK when you want a maintained Python client aligned with the ProofAgent™ platform and API.

Documentation and examples

Resource	Description
Documentation portal	Main product and SDK documentation
docs/python-sdk-guide.md	Python SDK guide
docs/quickstart.md	Quickstart snippets
examples/	Runnable examples

Build docs locally:

make docs-serve

Configuration

Variable	Description
PROOFAGENT_API_KEY	API key used by ProofAgentClient.from_env()
PROOFAGENT_BASE_URL	API base URL. Defaults to https://api.proofagent.ai

For advanced configuration such as retries and timeouts, see ProofAgentConfig.

Package layout

src/proofagent/ — main SDK package

Module	Role
proof_agent.py	ProofAgent facade (evaluate_sync, reasoning defaults)
tested_agent.py	TestedAgent (JSON + handler or endpoint)
client.py	ProofAgentClient (evaluate, evaluate_logs, REST)
evaluation.py	EvaluationResult (score, label) and helpers
project_support.py	Log-Based project checks (assert_project_supports_logs)
config.py	Configuration handling
exceptions.py	SDK exceptions
types.py	Shared SDK types
cli.py	CLI entrypoint for the proofagent command

Runtime requirements: Python 3.10+, httpx for async HTTP.

License

See the LICENSE file for details.

Support

• Website: https://www.proofagent.ai
• Documentation: https://www.proofagent.ai/docs
• GitHub Issues: https://github.com/ProofAgent-ai/proofagent-sdk/issues