VeilPiercer

Per-step tracing for local LLM pipelines. Zero cloud. Zero config.

When your AI agent breaks, the failure didn't happen at the error. It happened 3–5 steps earlier when agents silently diverged. VeilPiercer tells you exactly which step went wrong, what it read, and what it produced.

Works with Ollama, LangChain, LangGraph, CrewAI — 100% offline, no API key needed.

---

Install

pip install veilpiercer

---

Quickstart — 60 seconds

LangChain / LangGraph

from veilpiercer import VeilPiercerCallback

cb = VeilPiercerCallback(session_id="my-run-1")

# Pass to any LangChain chain, agent, or LLM
result = chain.invoke({"input": "..."}, config={"callbacks": [cb]})

Ollama (monkey-patch, 1 line)

from veilpiercer import enable_ollama_tracing
enable_ollama_tracing()  # all subsequent ollama.chat() calls are traced

Manual logging

from veilpiercer.logger import SessionLogger

with SessionLogger.new("my-session") as s:
    s.log_step("prompt", "What is the capital of France?", model="llama3")
    s.log_step("response", "Paris.", model="llama3")

---

CLI

# List all captured sessions
veilpiercer sessions

# Open the visual diff UI in your browser
veilpiercer dashboard

# Compare two sessions side-by-side
veilpiercer diff <session_a> <session_b>

# Run a demo (no agent needed)
veilpiercer demo

# Check version
veilpiercer --version

---

What It Captures

Every session records:

Event	What's stored
LLM calls	prompt, output, model, token count, latency
Tool calls	tool name, input, output
Chain start/end	session metadata, duration
Agent steps	thought, action, observation
Retrievals	query, documents returned
Errors	exception type, traceback step

All stored locally in SQLite (nexus_mind.db). Nothing leaves your machine.

---

Visual Diff UI

veilpiercer demo      # generates two sample sessions
veilpiercer dashboard # opens the diff UI in your browser

Select any two sessions and see exactly where they diverged — which step produced different output, what each version of the agent saw.

---

How It Works

Your Agent
    ↓ (2 lines of code)
VeilPiercerCallback
    ↓ captures every step
nexus_mind.db (local SQLite)
    ↓
veilpiercer dashboard (visual diff)

No backend. No cloud. No telemetry. Sessions are stored in nexus_mind.db in your working directory.

---

Privacy

• 100% local — SQLite only, nothing is sent anywhere
• Works fully offline — no API key, no account required
• Compatible with air-gapped environments

---

Compatibility

Framework	Support
LangChain / LCEL	✅ Full (LLMs, tools, chains, agents, retrievers)
Ollama	✅ 1-line hook
LangGraph	✅ Via LangChain callbacks
CrewAI	✅ Via LangChain callbacks
AutoGPT	✅ Via LangChain callbacks
Raw Python	✅ SessionLogger directly

---

Hosted Dashboard (optional)

pip install veilpiercer is free forever for local use.

For cloud dashboard, team sharing, and real-time alerts, get lifetime access at veil-piercer.com — $197 one-time, no subscription.

---

License

MIT © VeilPiercer