Route a dataset -> writes results.jsonl + summary.json + an index.html dashboard

<div align="center">

![pip install tryaii](https://img.shields.io/badge/pip-tryaii-2563eb?style=flat-square&logo=pypi&logoColor=white)
![npm install tryaii](https://img.shields.io/badge/npm-tryaii-ef4444?style=flat-square&logo=npm&logoColor=white)
![python 3.9+](https://img.shields.io/badge/python-3.9%2B-2563eb?style=flat-square&logo=python&logoColor=white)
![node 18+](https://img.shields.io/badge/node-18%2B-ef4444?style=flat-square&logo=node.js&logoColor=white)
![license Apache 2.0](https://img.shields.io/badge/license-Apache%202.0-555?style=flat-square)

</div>

```text
████████╗██████╗ ██╗   ██╗ █████╗ ██╗██╗
╚══██╔══╝██╔══██╗╚██╗ ██╔╝██╔══██╗██║██║   ▸ Diff Routing Engine
   ██║   ██████╔╝ ╚████╔╝ ███████║██║██║   ▸ semantic, prompt-aware LLM routing
   ██║   ██╔══██╗  ╚██╔╝  ██╔══██║██║██║   ▸ ranks LLMs by benchmark × cost × speed
   ██║   ██║  ██║   ██║   ██║  ██║██║██║   ▸ local embeddings — zero API keys to route
   ╚═╝   ╚═╝  ╚═╝   ╚═╝   ╚═╝  ╚═╝╚═╝╚═╝
```

> **TryAii** reads your prompt, figures out *what kind of task it is* using local
> embeddings, and routes it to the best LLM for the job — balancing benchmark quality,
> price, and latency the way **you** tell it to. The same engine ships as a Python
> package and a Node/TypeScript package, with one matching `tryaii` CLI.
>
> _(The wordmark above animates in a blue→red gradient when you run the CLI in a real terminal.)_

**Rather watch the race than predict it?** The same model catalog powers
[**tryaii.com**](https://tryaii.com) — a side-by-side playground where one prompt
fans out to OpenAI, Anthropic, Google, xAI, DeepSeek, and Mistral at once and
streams back with tokens, latency, and per-call cost inline. `tryaii` picks the
winner before you spend; tryaii.com lets you watch them run.

## Demo

https://github.com/user-attachments/assets/b9f4d361-5c63-4138-ab13-f589e959798d

---

## Install

```bash
pip install tryaii        # Python 3.9+
npm install tryaii        # Node 18+
```

Both install a `tryaii` command on your `PATH`. Routing runs **fully locally** —
embeddings are computed on-device (`sentence-transformers` on Python, ONNX MiniLM via
`@xenova/transformers` on Node), so no API key is needed just to rank models. An
OpenRouter key is only required if you want the SDK to *call* the chosen model for you.

Out of the box, `tryaii` routes on a **starter catalog** of 45 well-known models and
16 benchmarks that ships inside the package — no account, no network. Log in (free,
with your tryaii.com Google account) to route on the **full catalog of 322 models**:

```bash
tryaii login     # device sign-in: open the printed URL in any browser, approve the code
tryaii whoami    # shows the account and the downloaded catalog release
tryaii logout    # removes the credentials and the downloaded catalog
```

After login the full catalog downloads automatically, is refreshed at most once a day,
cached locally, and checked against an Ed25519 signature before it is used
([docs](docs/cli/login.md)).

Building a real multi-model product? See the local-first
[design-partner program](docs/design-partners.md).

## Examples

The most common use is `tryaii eval` — route a whole dataset under a budget:

```bash
# Route a dataset -> writes results.jsonl + summary.json + an index.html dashboard
# (a ready-made 1000-prompt dataset ships at examples/prompts.json)
tryaii eval examples/prompts.json --output results/run

# Spend at most $0.50 total; invest more in the harder prompts (default: intrinsic difficulty)
tryaii eval examples/prompts.json --max-price=0.50 --output-tokens=2000

# Gauge difficulty from model disagreement instead, and push budget harder toward hard prompts
tryaii eval examples/prompts.json --max-price=0.50 --difficulty-source=capability --difficulty-gamma=3

# Shrink answers to fit a tight budget instead of failing
tryaii eval examples/prompts.json --max-price=0.10 --output-tokens=2000 --budget-mode=fit-output
```

Or rank models for a single prompt:

```bash
tryaii route "Debug this memory leak in my Node.js app" --quality=5 --cost=1 --speed=2
tryaii models --provider anthropic     # inspect the model catalog
```

Or lint a request for prompt-cache readiness before sending it — 18 dynamic-content
detectors, per-model token floors for 7 providers, and predicted HIT/MISS across
request sequences ([docs](docs/cli/cachelint.md)):

```bash
tryaii cachelint request.json                  # verdict-first report; warns, never blocks
cat prompt.txt | tryaii cachelint - --provider anthropic --model claude-fable-5
```

The same engine runs inside the SDK clients: `cache_lint="warn"` (Python) /
`cacheLint: 'warn'` (Node) lints every outgoing chat request at the moment it
leaves and verifies predictions against the provider's `usage` response —
warn-only, fail-open, once per prompt shape ([docs](docs/sdk/client/cache-lint.md)).

Or diagnose a whole codebase's LLM call sites — your coding agent finds the calls
and writes an inventory; `tryaii` runs four dete

... (truncated)