AI Resource Radar is a local-first tracker for free AI tiers, GPU compute, grants, and prices. It keeps the source and verification time beside each result, so the answer is practical: what is free, how much you get, when it resets, what the restrictions are, and how to claim it.

The default collection pipeline is deterministic: no AI, API key, cookie, or account data is required. AI is used only by the optional daily poster.

Start with uvx

Try the keyless collector without creating a checkout or virtual environment:

uvx ai-resource-radar start --open

For the hosted snapshot, use Live Radar or fetch the documented public data manifest. The public site is an aggregate view; always follow an official source before relying on an offer.

What you get

Free tiers and AI prices change frequently, while ordinary link lists quickly become stale. This project turns public source material into a small, explainable local database:

Quick start

Requires Python 3.11 or newer. Install from PyPI in an isolated environment:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install ai-resource-radar

ai-radar refresh
ai-radar dashboard --open

The dashboard is available only on 127.0.0.1:18766.

On macOS, install the dashboard, menu bar helper, and 08:00 daily job:

ai-radar service install
ai-radar service status

To uninstall the services without deleting the database:

ai-radar service uninstall

Platform support

Windows is not tested yet. Linux CI verifies the deterministic core and dashboard; macOS CI also compiles and tests the Vision OCR and menu bar helpers.

What it tracks

The built-in adapters currently cover 23 sources:

Community sources can discover candidates but cannot upgrade an offer to “officially verified”. Each HTTPS source is allow-listed, limited to 16 MB, isolated on failure, and supports ETag/Last-Modified caching.

Explainable ranking

The radar deliberately avoids an opaque score:

Within a tier, results are ordered by mainland availability, estimated value, and recent changes. The dashboard shows the reasons instead of hiding them inside a number.

Optional extensions

Daily poster

The image above is an explicitly labelled example. Always check the current official source before using an offer.

The poster model draws the complete image; the application never overlays or rewrites its text. The five facts are selected by deterministic code. macOS Vision then checks every required title, provider, value, action, statistic, update time, and unexpected number.

ai-radar poster models
ai-radar poster configure --provider openclaw --model zai/cogview-3-flash --disable
ai-radar poster benchmark status
ai-radar poster benchmark              # up to three cases per day
ai-radar poster benchmark review --approve
ai-radar poster configure --provider openclaw --model zai/cogview-3-flash --enable
ai-radar poster latest

The free zai/cogview-3-flash model is called through OpenClaw at its official 864×1152 portrait size and then proportionally resized to a final 1080×1440 WebP. Benchmark and daily generation share a hard limit of three image calls per day. Formal eligibility requires all six OCR and numeric allow-list cases across at least two calendar days, followed by a human check for ghosting, cropping, and layout errors. Until then the poster stays disabled; the system does not fall back to a paid OpenAI image call and radar refreshes continue normally.

AI efficiency tips

Official guidance and manually imported articles remain candidates until a human approves them. Approval is limited to the marked managed block in an AGENTS.md, creates a private backup, and is auditable and reversible. See docs/TIPS.md for the safety model and ai-radar tips --help for the CLI.

How it works

flowchart LR
    A[Allow-listed public sources] --> B[Source-specific deterministic parsers]
    B --> C[Normalized SQLite schema]
    C --> D[Explainable ranking and change detection]
    D --> E[Dashboard, CLI, and local notifications]
    D --> F[Deterministic five-fact selection]
    F --> G[Optional registered image model]
    G --> H[Local Vision OCR]
    H -->|Pass| E
    H -->|Fail, max 3/day| I[Discard candidate and keep last valid poster]

One failed source never clears data from other sources. A missing offer is removed only after two successful parses both confirm its absence. Parser drift keeps the last trusted value and marks the source for verification.

Useful commands

# Refresh all due sources, or bypass cadence
ai-radar refresh
ai-radar refresh --force

# Browse verified, no-card resources
ai-radar list --verified-only --no-card
ai-radar list --kind gpu --no-card

# Review recent changes
ai-radar changes --days 30

# Run the complete daily workflow
ai-radar daily

# Diagnose the database, source freshness, helpers, and services
ai-radar doctor
ai-radar doctor --json

Run ai-radar <command> --help for every filter and option.

Data, privacy, and storage

• SQLite schema v7 uses file mode 0600; no secrets, cookies, or account data are stored.
• Tips retain only bounded summaries and evidence. Approval updates only the marked AGENTS.md managed block and creates a private backup under ~/.codex/backups/ai-tips/.
• Full fetched pages are parsed in memory and are not archived.
• Fetch logs are retained for 90 days; ordinary changes and delivered notifications for 365 days.
• Important free-tier changes and unread notifications are retained.
• Posters are retained for 90 days; failed candidates are deleted immediately.
• Periodic cleanup and threshold-based VACUUM prevent unbounded growth.
• The dashboard accepts only loopback Host/Origin requests and serves no remote assets.

See Architecture and Security for details. Existing users should also read the v0.2 migration guide; complete removal is documented in Uninstall.

Development

git clone https://github.com/xxy12341/ai-resource-radar.git
cd ai-resource-radar
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

python -m unittest discover -s tests -p 'test_*.py'
node --check src/ai_resource_radar/web/ai-resources.js

Contributions are welcome. Start with CONTRIBUTING.md, or open an issue with the official source URL and the policy or price that needs attention.

License

MIT