AIMI

AI Model Inventory

An evidence-first catalogue for discovering, comparing, configuring, and monitoring AI models across providers and agent harnesses.

Built and maintained by Aubrey Zemba
Follow me on X / Twitter

---

What AIMI does

AIMI answers a practical question:

Which model can I use, through which provider, in which harness, with what limits, at what cost, and what evidence supports that answer?

It keeps canonical models separate from provider routes, records evidence at field level, tracks access semantics such as genuinely free and subscription-included, and monitors official provider endpoints for changes.

Installation (agent-first, macOS, Windows, and Linux)

A human does not need to install AIMI manually. An agent should clone or copy this repository, then run the portable installer from the repository root:

python install.py

Use py install.py on Windows when python is not the registered command. The installer:

1. Requires Python 3.11 or newer.
2. Creates the private local aimi.db from schema_v2.sql when it does not exist.
3. Installs the repository's bundled skill to ~/.agents/skills/model-catalogue/ using the platform's home directory.
4. Prints the exact database, skill, and CLI paths after completion.

The installer has no third-party Python dependencies. It uses only the Python standard library and SQLite. API keys are optional for local inspection; provider refreshes and health checks require the relevant environment variables.

Verify the installation:

python aimi summary
python validate_catalogue.py

On Windows, use py instead of python if required. The Python CLI is the portable entry point on every operating system. The optional skills/scripts/catalogue wrapper is provided for POSIX shells only; Windows agents should invoke python aimi and the Python maintenance scripts directly.

Platform notes

• Windows: Python 3.11+, SQLite via Python, and PowerShell or another agent shell are sufficient for the core CLI, database, validation, export, and catalogue scripts. Windows-specific harness paths are detected when available; missing harnesses are reported rather than invented.
• macOS/Linux: Python 3.11+ is sufficient for the core CLI. skills/scripts/catalogue can be used from a POSIX shell. hyperfine, zsh, and Hermes are optional integrations used only by the relevant monitoring workflows.
• Provider access: install the provider's own CLI or credentials only when you want to scan or test that provider. AIMI does not silently install agent harnesses or create API keys.

Quick start

python aimi summary
python aimi where deepseek-v4-flash
python aimi recommend --task coding --free

The main CLI is called aimi. It can search routes, compare providers, inspect harness configuration, show free-model health, and manage Pi model ordering.

Project layout

Path	Purpose
aimi	CLI for search, recommendations, configuration, and Pi ordering
aimi.db	Private authoritative SQLite database, kept out of Git
free_models.db	Compatibility symlink to aimi.db
monitor_endpoints.py	Polls official provider model endpoints and records changes
free_model_health.py	Runs bounded exact-OK health checks against eligible no-charge routes
scan_local_harnesses.py	Scans local harness configuration and availability
refresh_catalog.py	Refreshes provider routes from official endpoints
validate_catalogue.py	Checks integrity, evidence, pricing, ordering, and secret rules
export_sanitized.py	Creates a public database export with private state removed
schema_v2.sql	Core model, provider, evidence, harness, and monitoring schema
evidence/ and snapshots/	Local raw evidence and endpoint captures, never committed

The canonical cross-agent skill remains at:

$HOME/.agents/skills/model-catalogue/SKILL.md

AIMI is the project and CLI name. model-catalogue is the internal skill name used by the agent tooling.

CLI examples

Models and routes

./aimi summary
./aimi doctor
./aimi latest --limit 20
./aimi free
./aimi free --provider openrouter
./aimi where deepseek-v4-flash
./aimi where deepseek-v4-flash --refresh-harnesses
./aimi recommend --task coding --free

aimi where returns every matching provider route, its access semantics, known limits, harness matches, freshness, and the latest test outcome when one exists.

Harness inventory

./aimi harnesses
./aimi harness-models pi
./aimi harness-models droid
./aimi harness-models codex-cli --kind configured
./aimi harness-models codex-cli --kind available

The scanner reads the local sources used by each harness, including Pi, Droid, OpenCode, Codex CLI, Cline, Aside, Antigravity, and Mistral Vibe where present.

Pi model ordering

./aimi order-diff pi
./aimi pi-fragment openrouter 'poolside/laguna-s-2.1:free'
./aimi pi-register openrouter 'poolside/laguna-s-2.1:free'
./aimi pi-select openrouter 'poolside/laguna-s-2.1:free' --position 3

Writes preview by default. Applying a change creates a timestamped backup, then the local configuration is rescanned and validated.

Subscriptions and harness-specific access

./aimi subscriptions
./aimi subscriptions --mine
./aimi subscription opencode-go
./aimi warp-models
./aimi warp-models --custom

Subscription access is tracked separately from genuinely free API access. OpenCode Go and OpenCode Zen are separate products with separate endpoints. Warp BYOK and custom inference endpoints are recorded separately from Warp-hosted inference.

Evidence and pricing rules

AIMI treats these as different access categories:

Offer type	Meaning
genuine_zero_price	Official source proves that the route has no charge
free_tier_quota	Official developer or evaluation quota with explicit limits
temporary_free_window	Free for a window whose dates may be finite or unknown
subscription_included	Available through a subscription, not classified as free
paid	Paid API usage
unknown	Not enough evidence to classify the route

A working API key, open weights, free chat access, or free credits does not prove that an API route is free. Missing capability or context data means unverified, not unsupported.

AIMI keeps announcement, general availability, API availability, model-card, weights-release, endpoint-first-seen, and free-window events separate. Claims point to immutable evidence captures rather than silently replacing conflicting sources.

Health monitoring

Only routes currently verified as no-charge are eligible for free-model health probes. The probe contract is exact:

Reply with exactly OK

Results use three states:

• Green: exact OK returned.
• Orange: rate limited, so availability is inconclusive.
• Red: failed, unauthorized, timed out, or returned something other than exact OK.

The database keeps one current status per route and a bounded daily history. NVIDIA NIM developer-tier checks run separately from the regular OpenRouter and OpenCode cycle.

hyperfine --runs 1 --warmup 0 --show-output \
  -n 'free-model-health' \
  "zsh -lc 'source ~/.zshrc >/dev/null 2>&1; cd \"$HOME/AZ Labs/2 - Testing/AIMI\"; python3 free_model_health.py --workers 4 --timeout 45 --retention-days 30'"

./aimi free-health-summary
./aimi free-health
./aimi free-health --failures-only

Monitoring official endpoints

The endpoint monitor covers OpenRouter, OpenCode Zen, OpenCode Go, NVIDIA NIM, DeepSeek, Mistral, OpenAI, Gemini, Cloudflare Workers AI, and Ollama Cloud. It saves timestamped captures, hashes evidence, normalizes volatile fields, records additions and removals, and updates free-offer windows without storing secrets.

The Hermes job is:

model-catalogue-endpoint-monitor (454c1f94d5fe)

Inspect it with:

hermes cron runs 454c1f94d5fe

Public export

The private database, raw endpoint captures, local harness state, personal rankings, account identifiers, and credential inventory stay local. To produce a sanitized database for review or publication:

./export_sanitized.py

The exporter removes private state, runs SQLite integrity validation, and performs a fail-closed secret-pattern scan before writing the public export under dist/.

Development checks

./scan_local_harnesses.py
./validate_catalogue.py
./export_sanitized.py

The latest machine-readable validation result is stored in validation-report.json.

Privacy

AIMI is designed around a private local catalogue with a safe public export. API-key values are not stored in the database. Raw evidence, snapshots, local configuration inventories, and the private database are excluded from the Git repository.

Roadmap

• Link more provider routes to canonical model identities and aliases.
• Add reviewed capability and benchmark evidence for newer models.
• Expand safe runtime tests for tools, reasoning, images, streaming, and structured output.
• Add task-specific Pi order profiles and usage-based recommendations.
• Improve notifications for new free windows, removals, and deprecations.

---

Made by Aubrey Zemba · @TH33_ORACL3 on X