Local-first audit & trust console for AI agents — see what an agent did, why, with what evidence, and where it went wrong.

pip install peaky-peek-server && peaky-peek --open

Local-first, open-source agent audit debugger. Every run answers five operator questions — what happened, why, with what evidence, with what result, and where it failed — with deterministic claim verification and an explainable trust score, all on your machine.

Why Peaky Peek?

Traditional observability tools weren't built for agent-native debugging, and they don't answer the question operators actually care about: can I trust what this agent did?

Peaky Peek is a black-box recorder + reasoning audit console for AI agents. It captures the causal chain behind every action, then reframes each run as an audit record that answers five questions:

1. What happened? — the exact sequence of tool calls, model calls, decisions, retries, and outputs
2. Why? — the stated rationale, alternatives considered, confidence, and trigger for each important step
3. With what evidence? — the inputs used: user input, retrieved docs, tool results, prompt fragments
4. With what result? — success/failure, returned data, state changes, downstream effects
5. Where did it fail? — the first bad decision, ignored evidence, weak tool data, contradictions, plan drift, and the downstream damage

Every claim is classified deterministically as verified · partially verified · contradicted · unsupported · unverified, and each session gets an explainable trust score so a human can audit a run without guessing.

Quick Start

Option 1: Decorator (simplest)

pip install peaky-peek-server
peaky-peek --open   # launches API + UI at http://localhost:8000

from agent_debugger_sdk import trace

@trace
async def my_agent(prompt: str) -> str:
    # Your agent logic here — traces are captured automatically
    return await llm_call(prompt)

Option 2: Context Manager

from agent_debugger_sdk import trace_session

async with trace_session("weather_agent") as ctx:
    await ctx.record_decision(
        reasoning="User asked for weather",
        confidence=0.9,
        chosen_action="call_weather_api",
        evidence=[{"source": "user_input", "content": "What's the weather?"}],
    )
    await ctx.record_tool_call("weather_api", {"city": "Seattle"})
    await ctx.record_tool_result("weather_api", result={"temp": 52, "forecast": "rain"})

Option 3: Zero-Config Auto-Patch (no code changes)

# Set env var, then run your agent normally
PEAKY_PEEK_AUTO_PATCH=true python my_agent.py

Works with PydanticAI, LangChain, OpenAI SDK, CrewAI, AutoGen, LlamaIndex, and Anthropic — no imports or decorators needed.

Framework Integrations

PydanticAI

from pydantic_ai import Agent
from agent_debugger_sdk import init
from agent_debugger_sdk.adapters import PydanticAIAdapter

init()

agent = Agent("openai:gpt-4o")
adapter = PydanticAIAdapter(agent, agent_name="support_agent")

LangChain

from agent_debugger_sdk import init
from agent_debugger_sdk.adapters import LangChainTracingHandler

init()

handler = LangChainTracingHandler(session_id="my-session")
# Pass handler to your LangChain agent's callbacks

OpenAI SDK

No code needed — just set the environment variable:

PEAKY_PEEK_AUTO_PATCH=true python my_openai_agent.py

Or use the simplified decorator:

from agent_debugger_sdk import trace

@trace(name="openai_agent", framework="openai")
async def my_agent(prompt: str) -> str:
    client = openai.AsyncOpenAI()
    response = await client.chat.completions.create(
        model="gpt-4o", messages=[{"role": "user", "content": prompt}]
    )
    return response.choices[0].message.content

Auto-Patch (Any Framework)

import agent_debugger_sdk.auto_patch  # activates on import when PEAKY_PEEK_AUTO_PATCH is set

# Now run your agent normally — all LLM calls are traced automatically

Features

Agent Audit & Trust Console

Peaky Peek's defining capability: every session produces an audit report that turns a trace into evidence. Open the Audit panel on any session to see:

• Trust header — an explainable score (low / medium / high) with its components: evidence coverage, verification rate, policy compliance, recovery rate, failure severity, and contradiction count.
• The five-question view — What happened · Why · Evidence used · Outcome · Where it failed, in one grid.
• Verification badges — every decision is tagged verified, partially_verified, contradicted, unsupported, or unverified, with the basis (tool result, user input, retrieved doc, or none).
• Where-it-failed — the first bad decision, localized failure root-cause suspects, and the causal path to each failure.
• Risk signals — deterministic detections: unsupported claims, missing evidence, contradictions, repeated failed strategies, plan drift, policy violations, weak evidence.

Every row is clickable and jumps to the underlying event. The same report is available as JSON at GET /api/sessions/{id}/audit. Deterministic only — no opaque "AI insights," every number is derivable from captured fields.

See the Audit & Trust guide for the data model, an example audited session, and an example failure report.

Decision Tree Visualization

Navigate agent reasoning as an interactive tree. Click nodes to inspect events, zoom to explore complex flows, and trace the causal chain from policy to tool call to safety check.

Checkpoint Replay

Time-travel through agent execution with checkpoint-aware playback. Play, pause, step, and seek to any point in the trace. Checkpoints are ranked by restore value so you jump to the most useful state.

Trace Search

Find specific events across all sessions. Search by keyword, filter by event type, and jump directly to results.

Failure Clustering & Multi-Agent Coordination

Adaptive analysis groups similar failures. Inspect planner/critic debates, speaker topology, and prompt policy parameters across multi-agent systems.

Session Comparison

Compare two agent runs side-by-side. See diffs in turn count, speaker topology, policies, stance shifts, and grounded decisions.

Privacy & Security

• Local-first by default — no external telemetry, no data leaves your machine
• Zero-config auto-patching — no credentials or API keys needed for local debugging
• Optional redaction pipeline — prompts, payloads, PII regex
• API key authentication — bcrypt hashing
• GDPR/HIPAA friendly — SQLite storage, no cloud dependency

Deployment

pip (recommended)

pip install peaky-peek-server
peaky-peek --open

Docker

docker build -t peaky-peek .
docker run -p 8000:8000 -v ./traces:/app/traces peaky-peek

Development

git clone https://github.com/acailic/agent_debugger
cd agent_debugger
pip install -e ".[dev]"
pip install fastapi "uvicorn[standard]" "sqlalchemy[asyncio]" aiosqlite alembic aiofiles bcrypt
python3 -m pytest -q
cd frontend && npm install && npm run build

Architecture

System Overview

flowchart TB
    classDef layer fill:#0f172a,stroke:#334155,color:#e2e8f0,stroke-width:2px
    classDef ext fill:none,stroke:#94a3b8,stroke-dasharray:6 3,color:#94a3b8

    AGENT("🤖  Your Agent Code"):::ext

    subgraph RUNTIME[" "]
        direction TB
        SDK["<b>🔌  SDK Layer</b><br/><small>Instrument & capture</small><br/><sub>@trace · TraceContext · Auto-Patch · Adapters</sub>"]:::layer
        INTEL["<b>🧠  Intelligence</b><br/><small>Detect, remember, alert</small><br/><sub>Event Buffer · Pattern Detector · Failure Memory · Replay Engine</sub>"]:::layer
    end

    subgraph SERVER[" "]
        direction TB
        API["<b>🌐  API Server</b><br/><small>FastAPI + SSE</small><br/><sub>11 routers: sessions · traces · replay · search · analytics · compare</sub>"]:::layer
        STORE["<b>💾  Storage</b><br/><small>SQLite WAL · async</small><br/><sub>Events · Checkpoints · Analytics · Embeddings</sub>"]:::layer
    end

    UI["<b>🖥️  Frontend</b><br/><small>React · TypeScript · Vite</small><br/><sub>8 panels: decision tree · timeline · tools · replay · search · analytics · compare · live</sub>"]:::layer

    AGENT ==>|"decorate"| SDK
    SDK ==>|"emit"| INTEL
    INTEL -->|"persist"| STORE
    SDK -.->|"ingest"| API
    API <-->|"query"| STORE
    API ==>|"SSE stream"| UI
    INTEL -.->|"replay"| API

Layer Detail

flowchart LR
    classDef sdk fill:#4f46e5,stroke:#3730a3,color:#fff,stroke-width:2px
    classDef intel fill:#dc2626,stroke:#b91c1c,color:#fff,stroke-width:2px
    classDef api fill:#059669,stroke:#047857,color:#fff,stroke-width:2px
    classDef store fill:#b45309,stroke:#92400e,color:#fff,stroke-width:2px
    classDef ui fill:#7c3aed,stroke:#6d28d9,color:#fff,stroke-width:2px

    subgraph SDK[" 🔌  SDK "]
        direction TB
        DEC["@trace decorator"]:::sdk
        CTX["TraceContext"]:::sdk
        AP["Auto-Patch"]:::sdk
        AD["Framework Adapters"]:::sdk
    end

    subgraph INT[" 🧠  Intelligence "]
        direction TB
        BUF["Event Buffer"]:::intel
        PAT["Pattern Detector"]:::intel
        FMEM["Failure Memory"]:::intel
        ALERT["Alert Engine"]:::intel
        RPLAY["Replay Engine"]:::intel
    end

    subgraph APIL[" 🌐  API "]
        direction TB
        R1["Sessions · Traces"]:::api
        R2["Replay · Search"]:::api
        R3["Analytics · Compare"]:::api
        SSE["SSE Stream"]:::api
    end

    subgraph STO[" 💾  Storage "]
        direction TB
        DB[("SQLite WAL")]:::store
        S1["Events · Checkpoints"]:::store
        S2["Analytics Aggregations"]:::store
    end

    subgraph UIF[" 🖥️  Frontend "]
        direction TB
        DT["Decision Tree"]:::ui
        TL["Trace Timeline"]:::ui
        TI["Tool Inspector"]:::ui
        RP["Session Replay"]:::ui
        SE["Cross-session Search"]:::ui
        AN["Analytics Dashboard"]:::ui
    end

    DEC & CTX --> BUF
    AP & AD --> BUF
    BUF --> PAT & FMEM & ALERT
    BUF --> S1
    S1 --> DB
    DB --> S2
    R1 & R2 & R3 <--> S1
    RPLAY --> R2
    SSE --> DT & TL & RP

See ARCHITECTURE.md for full module breakdown.

Project Status

• Agent audit & trust — deterministic 5-questions report, claim verification, risk signals, explainable trust score (API + Audit UI panel)
• Core debugger — local path end-to-end, stable
• SDK — @trace, trace_session(), auto-patch for 7 frameworks
• API — 12 routers: sessions, traces, replay, search, analytics, cost, comparison, audit
• Frontend — 9 specialized panels (decision tree, replay, checkpoints, search, audit)
• Tests — 2900+ passing, CI on Python 3.10/3.11/3.12

Scientific Foundations

Peaky Peek is informed by research on agent debugging, causal tracing, failure analysis, and adaptive replay. See paper notes for design takeaways from each.

• AgentTrace: Causal Graph Tracing for Root Cause Analysis
• XAI for Coding Agent Failures
• FailureMem: Failure-Aware Autonomous Software Repair
• MSSR: Memory-Aware Adaptive Replay
• Learning When to Act or Refuse
• Policy-Parameterized Prompts
• CXReasonAgent: Evidence-Grounded Diagnostic Reasoning
• NeuroSkill: Proactive Real-Time Agentic System
• REST: Receding Horizon Explorative Steiner Tree
• Towards a Neural Debugger for Python

Documentation

• 5-Minute Getting Started
• Audit & Trust Guide
• Integration Guide
• SDK README
• Architecture Overview
• Progress Tracker

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

License

MIT