Chowkidar

Chowkidar (चौकीदार) means watch guard in Hindi — a local sentinel that keeps watch over the LLM models in your codebase so deprecated models never take production down.

Chowkidar is a secure, local-first LLM model deprecation watchdog. It scans your codebase and configuration files for LLM model references, cross-references them with a locally cached deprecation database, and alerts you before models sunset.

By default everything runs on your machine with zero data exfiltration. Remote cloud advisory classification is strictly opt-in for CI/CD environments where a local SLM is unavailable.

Architecture

flowchart TD
    subgraph input [Input]
        scan[Scan: .env, YAML, TOML, source code]
        cloud[Cloud secrets: AWS, GCP, Azure, Vercel, K8s]
    end

    subgraph registry [Local Registry]
        sync[Provider sync: OpenAI, Anthropic, Google, Mistral]
        db[(SQLite registry.db)]
    end

    subgraph advisory [Hybrid Advisory Engine]
        cache{advisory_cache.json hit?}
        tier1{Cloud advisory enabled?}
        tier2{Local SLM available?}
        remote[Remote connectors: OpenAI / Anthropic / Gemini / Mistral]
        slm[Local SLM: Ollama]
        heuristics[Deterministic heuristics]
    end

    subgraph validation [Validation Envelope]
        capGuard[Capability regression guard]
        localReg[Local registry successor mapping]
    end

    subgraph output [Output]
        alerts[Desktop / Slack / Discord alerts]
        rules[IDE rules: Cursor, Claude, Copilot, Windsurf]
        mcp[MCP server]
        report[HTML / Markdown / JSON reports]
    end

    scan --> db
    cloud --> scan
    sync --> db
    scan --> cache
    cache -->|miss| tier1
    tier1 -->|yes| remote
    tier1 -->|no| tier2
    tier2 -->|yes| slm
    tier2 -->|no| heuristics
    remote --> capGuard
    slm --> capGuard
    heuristics --> capGuard
    capGuard --> localReg
    localReg --> alerts
    localReg --> rules
    localReg --> mcp
    localReg --> report

Flow in plain terms: Chowkidar scans your project (and optionally cloud secret stores), syncs provider deprecation data into a local database, classifies each model's usage purpose through a three-tier engine (remote API → local SLM → heuristics), then applies deterministic successor selection and capability validation before alerting you or updating IDE rules.

Why Chowkidar?

In the rapidly evolving landscape of Large Language Models, providers (OpenAI, Anthropic, Google, Mistral) deprecate and sunset models at an unprecedented pace. When a model sunsets, your application's API calls fail — leading to immediate production outages, broken user experiences, and lost revenue.

Chowkidar acts as your automated, local-first sentinel to eliminate this risk.

What It Does For You

• Continuous Codebase & Cloud Auditing: Scans repositories, environment variables, GitHub Secrets, and cloud provider secret stores (AWS, GCP, Azure, Vercel, Kubernetes) for active LLM model references.
• Proactive Multi-Channel Alerting: Warns you via desktop notifications, Slack Block Kit, or Discord webhooks at critical thresholds (30, 15, 7, and 1 day) before a model sunsets.
• Hybrid Purpose Classification: Three-tier advisory engine — opt-in remote cloud connectors, local Ollama SLM, or offline heuristics — classifies how you use each model (coding, reasoning, embeddings, chat).
• Deterministic Successor Selection: Remote LLMs classify purpose only; successor models and capability validation always come from Chowkidar's local registry and capability regression guard.
• Automated, Safe Migrations: Previews and applies atomic updates to configuration files with automatic backups and rollback support.
• AI Editor Alignment: Generates passive workspace rules and runs an active MCP server so AI coding assistants (Cursor, Claude Code, Copilot, Windsurf) never write deprecated models.

The Value You Get

• Zero Production Outages: Migrate well in advance of provider deadlines.
• Privacy by Default: Runs locally and offline. Opt-in cloud advisory sends only minimized, redacted code snippets — never full files or secrets.
• No Capability Regressions: Compares context size, output tokens, vision, and tool-use support before recommending successors.
• Immediate FinOps Savings: Built-in cost-impact engine shows input/output token price deltas in percentage terms before you migrate.
• Zero-Config Developer Experience: chowkidar setup auto-configures your workspace, IDE rules, and background daemon in seconds.

Quick Start

# Install
pip install chowkidar

# Interactive project setup (config, DB, provider sync, initial scan, IDE rules)
chowkidar setup

# One-command demo: sync → scan → HTML report → open in browser
chowkidar showcase

Setup & Configuration Guide

For Human Developers

Local Installation & Project Initialization

pip install chowkidar
chowkidar setup

chowkidar setup creates .chowkidar/ (config + SQLite DB), syncs deprecation data, scans your project, and writes IDE rules.

Continuous Background Monitoring

chowkidar daemon

The daemon monitors watched projects and triggers alerts at 30, 15, 7, and 1 day before expiry.

Cloud Advisory (Opt-In, for CI/CD)

When Ollama is unavailable (e.g. GitHub Actions), enable remote purpose classification:

# .chowkidar/config.toml
cloud_advisory_enabled = true
cloud_advisory_provider = "openai"   # openai | anthropic | google | mistral
cloud_advisory_model = "gpt-4o-mini"

export CHOWKIDAR_OPENAI_API_KEY="sk-..."

Remote connectors batch-classify purposes with strict timeouts, secret redaction, and automatic fallback to local heuristics. They never select successor models.

Cloud & CI/CD Deployment

GitHub Actions:

name: Chowkidar Watchdog
on:
  push:
    branches: [main]
  schedule:
    - cron: '0 9 * * 1'
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run Chowkidar
        uses: bhav09/chowkidar@v0.9.5
        with:
          gate: true
          slack_webhook: ${{ secrets.SLACK_WEBHOOK }}

Docker:

docker run -d \
  -v $(pwd)/.chowkidar:/app/.chowkidar \
  -e CHOWKIDAR_SLACK_WEBHOOK="https://hooks.slack.com/services/..." \
  bhavishya/chowkidar:latest

For AI Agents & IDEs

Passive (workspace rules): chowkidar setup or chowkidar rules write generates rules for Cursor, Claude Code, Copilot, and Windsurf.

Active (MCP): Add to your IDE's MCP config:

{
  "mcpServers": {
    "chowkidar": {
      "command": "chowkidar",
      "args": ["mcp"]
    }
  }
}

Top 5 CLI Commands

1. chowkidar showcase

One-command demo run: syncs provider data, scans your project, generates an interactive HTML report, and opens it in your browser. Best way to see Chowkidar in action.

2. chowkidar setup

Project-scoped configuration, database initialization, provider sync, initial scan, and IDE rules generation.

3. chowkidar scan

Locates all LLM model references in your code and configuration files.

4. chowkidar check

Cross-references detected model strings against the deprecation registry and prints inline warnings.

5. chowkidar report

Generates comprehensive Markdown, JSON, or interactive HTML deprecation reports.

See COMMANDS.md for the complete command reference.

Security & Local Safety

• Privacy First: No code, secrets, or project paths leave your machine unless you explicitly enable cloud_advisory_enabled. When enabled, only minimized, redacted context snippets are sent.
• Safe Writes: File modifications require auto_update = true. Every update uses atomic writes and saves a .chowkidar.bak backup.
• Concurrent-Safe: System-level filelock prevents corruption from concurrent daemon/CLI writes.
• Capability Guard: Successor recommendations are validated against capability regressions before any auto-write is allowed.

License

MIT