clickproof

Persistent GUI behavioral facts for computer-use agents.

Quick Start · How It Works · CLI Reference · GitHub Action · vs. Alternatives · Claude/MCP · Contributing

---

Why

Computer-use agents navigate GUIs blindly. Every session restarts from zero — the agent re-discovers which button opens a dialog, which tab holds exports, which field triggers validation.

This is expensive. More importantly, it's fragile: apps change, and the agent's cached intuition from training is often wrong.

clickproof solves this by giving agents a persistent, confidence-scored memory of UI behavioral facts. Before a session starts, the agent loads what is known about the target app. Observations from every run update confidence scores. When an interface changes, scores decay and the agent adapts.

# Inject known facts into an agent's system prompt
clickproof query salesforce --min-score 0.7

---

How It Works

flowchart LR
    A[Agent records UIFact\napp · element · action → outcome] --> B[FactStore\nSQLite persistence]
    B --> C[FactObservation\nconfirmed or refuted]
    C --> D[FactScorer\nbase_ratio × staleness_decay × count_boost]
    D --> E[FactRetriever\nquery by app + min_score]
    E --> F[bootstrap_context\ntext for system prompt injection]

Core primitives:

• UIFact — an immutable, content-addressed record of app_name + app_version + element + action → outcome. ID = SHA-256[:16] of the key fields. Same element observed twice always produces the same ID.
• FactObservation — a confirmed/refuted signal from an agent run, linked to a UIFact.
• FactScorer — computes a confidence score from observation history: base_ratio × staleness_decay × count_boost.
• FactRetriever — queries facts by app and version, filtered by minimum score, and generates a text context string for agent injection.

---

Features

Feature	Details
Content-addressed facts	Same app/version/element/action always produces the same ID
Bayesian-style scoring	Score = base ratio × staleness decay × count boost
Staleness decay	Score decays exponentially at e^(-0.1 × staleness_days)
Offline / local-first	Single SQLite file, no server required
Agent context injection	bootstrap_context() returns a ready-to-inject text block
JSON output	Machine-readable output for downstream automation
Markdown output	Ready-to-paste format for issue comments and PRs
FastAPI REST server	/fact, /observe, /query, /facts, /bootstrap, /health endpoints
MCP server	Model Context Protocol tools for Claude and other MCP-compatible agents
166 tests	Comprehensive suite covering all layers with 87%+ branch coverage

---

Quick Start

pip install clickproof

Extras / Optional Dependencies

# FastAPI REST server (5 endpoints: /fact /observe /query /facts /bootstrap /health)
pip install 'clickproof[api]'
uvicorn clickproof.api:app --reload

# MCP server for Claude Desktop and other MCP-compatible agents
pip install 'clickproof[mcp]'

from clickproof import UIFact, FactObservation, FactStore, FactRetriever, FactScorer
import time

with FactStore("my_app.db") as store:
    # Record a UI behavioral fact
    fact = UIFact(
        app_name="salesforce",
        app_version="2025.11",
        element="export-csv-button",
        action="click",
        outcome="opens-download-dialog",
        context="reports-page",
    )
    store.add_fact(fact)

    # Record an observation confirming the fact
    obs = FactObservation(
        fact_id=fact.id,
        observed_at=time.time(),
        confirmed=True,
        agent_run_id="run_001",
    )
    store.add_observation(obs)

    # Retrieve facts for an app session
    retriever = FactRetriever(store, FactScorer())
    pairs = retriever.query(app_name="salesforce", min_score=0.5)
    for fact, score in pairs:
        print(f"[{score.score:.2f}] {fact.element} --{fact.action}--> {fact.outcome}")

    # Get a text block for agent context injection
    context = retriever.bootstrap_context("salesforce", "2025.11")
    print(context)

---

CLI Reference

clickproof [--db PATH] COMMAND [ARGS]

Commands:
  add     APP VERSION ELEMENT ACTION OUTCOME  Stage a UIFact
  observe FACT_ID --confirmed/--refuted       Record an observation
  query   APP [--version V] [--min-score F]   Retrieve scored facts (--format rich|json|markdown)
  log     [--app APP] [--json]                List all stored facts
  status                                      Show store info and stats
  decay   APP [--min-score F] [--format F]    Show score decay projections for an app
  export  APP [-o FILE] [--bootstrap]         Export facts as JSON (bootstrap pack optional)

Examples

# Add a fact
clickproof add salesforce 2025.11 export-csv-button click opens-download-dialog

# Confirm it from an agent run
clickproof observe <fact_id> --confirmed --run-id run_001

# Query with minimum score threshold
clickproof query salesforce --min-score 0.6

# Get JSON output for scripting
clickproof query salesforce --json | jq '.facts[].fact.element'

# Get Markdown output (ready to paste in issues / PRs)
clickproof query salesforce --format markdown

# Show score decay projections
clickproof decay salesforce --min-score 0.6

# Export facts to a file
clickproof export salesforce -o salesforce_facts.json

# Show store info
clickproof status

---

Formatters

clickproof ships three output formatters in clickproof.report (also importable from clickproof):

Function	Description
print_facts(pairs, console)	Rich-formatted console table
to_json(pairs)	JSON string — {"count": N, "facts": [...]}
to_markdown(pairs)	Markdown table — ready to paste in issue comments and PRs

from clickproof import FactRetriever, FactScorer, FactStore, to_markdown

with FactStore("my_app.db") as store:
    retriever = FactRetriever(store, FactScorer())
    pairs = retriever.query("salesforce", min_score=0.6)
    print(to_markdown(pairs))

---

GitHub Action

Add clickproof fact queries to any CI/CD workflow:

- uses: sandeep-alluru/clickproof@main
  with:
    app-name: salesforce
    app-version: "2025.11"
    db: clickproof.db
    min-score: "0.5"

---

vs. Alternatives

	clickproof	Plain cache	Vector store	Re-run
Confidence-based	✓	✗	partial	✗
Staleness decay	✓	✗	✗	N/A
Content-addressed	✓	✗	✗	N/A
Local-first	✓	✓	partial	✓
MCP native	✓	✗	partial	✗
Agent context injection	✓	manual	manual	N/A

---

Claude/MCP

clickproof ships a built-in MCP server. Add it to your Claude configuration:

{
  "mcpServers": {
    "clickproof-mcp": {
      "command": "clickproof-mcp",
      "env": { "CLICKPROOF_DB": "/path/to/clickproof.db" }
    }
  }
}

Available MCP tools: add_ui_fact, query_facts, bootstrap_context.

---

OpenAI / Tool Use

See tools/openai-tools.json for pre-built OpenAI function-calling tool definitions.

---

Case Studies

See how teams are using clickproof in production:

• Eliminating Session Startup Latency in Enterprise RPA with Persistent UI Facts
• Persistent CSS Selector Memory for High-Volume Web Data Extraction

---

Repository Tree

clickproof/
├── clickproof/
│   ├── __init__.py        Public API
│   ├── fact.py            UIFact + FactObservation data models
│   ├── scorer.py          FactScorer + FactScore
│   ├── store.py           SQLite-backed FactStore
│   ├── retriever.py       FactRetriever + bootstrap_context
│   ├── report.py          Rich / JSON / Markdown formatters
│   ├── cli.py             Click CLI
│   ├── api.py             FastAPI server
│   └── mcp_server.py      MCP server
├── tests/                 166 pytest tests
├── examples/
│   ├── demo.py                      Standalone walkthrough
│   ├── computer_use_agent.py        Computer-use agent integration
│   ├── multi_agent_shared_memory.py Multi-agent shared memory example
│   └── web_scraper_validation.py    Web scraper validation example
├── action.yml             GitHub Action
└── pyproject.toml

---

GitHub Topics

computer-use llm-agents agent-memory gui-automation behavioral-facts mcp llmops sqlite python

---

---

Stay Updated

Subscribe to The Silence Layer — weekly dispatches on production AI infrastructure, new releases, and the failure modes that production AI systems don't surface until it's too late.

Star History