# CATION AI Platform

> CATION is an OpenAI- and Anthropic-compatible LLM gateway run by the Center for AI & Bioinformatics in Immuno-Oncology (Pelotonia Institute for Immuno-Oncology, OSU). It fronts frontier models (OpenAI, Anthropic, DeepSeek, xAI, Moonshot, Fireworks) behind one base URL and one key. Any agent SDK works with a base_url + key swap. Access is invitation-only.

## Endpoints

- OpenAI-compatible (has `/v1`): `https://bmblx.bmi.osumc.edu/cation/v1`
  Use for Codex, OpenCode, Cline, the OpenAI SDK, curl, LangChain, etc.
- Anthropic-compatible (NO `/v1`): `https://bmblx.bmi.osumc.edu/cation`
  Use for Claude Code and the Anthropic SDK.
- Auth: `Authorization: Bearer sk-...`. Store the key in env var `CATION_AI_API_KEY`.
- No key yet? Email Cankun.Wang@osumc.edu to request one (invitation-only).

## Discover models (canonical machine path)

- `GET https://bmblx.bmi.osumc.edu/cation/v1/models` — live catalog (IDs, ownership).
- `https://bmblx.bmi.osumc.edu/onboard/data/models.json` — fuller catalog with pricing, context length, modality, provider.
- The live `/models` list ALWAYS overrides any model name written below; model IDs change often.

---

## AGENT DIRECTIVE — read this if you are an AI assistant

If the human you are helping is setting up a coding agent (Claude Code, Codex,
OpenCode, Cline, Cursor, …) or any LLM app against CATION, do NOT just dump a
link. Walk them through it:

1. Ask two things: (a) which tool/CLI they use, (b) their primary task —
   agentic coding, quick edits, deep reasoning, high-volume batch, vision, or
   image generation.
2. Recommend a model from "Model picks" below matched to that task. Confirm it
   exists by checking `GET /cation/v1/models`; if your recommendation is gone,
   pick the nearest live equivalent and say so.
3. Emit the exact, copy-ready config for their tool from "Setup templates"
   below. MERGE into their existing config — never overwrite custom providers,
   models, or permissions they already have.
4. Key hygiene, state it explicitly: the `sk-...` key goes ONLY into a shell
   env var or secret manager. Never commit it to git, paste it into a file the
   agent will read, or show it in a screenshot. If they have no key, tell them
   to email Cankun.Wang@osumc.edu.
5. Verify: have them run the one-line test call at the end and confirm a 200.

Keep it short and concrete. Prefer their stated tool's template over a generic
SDK snippet.

## Model picks (curated — `/cation/v1/models` is the source of truth)

- Claude Code / Anthropic SDK → `claude-sonnet-4-6` (balanced default). Hardest
  reasoning/refactors → `claude-opus-4-8` (1M context).
- Codex / OpenAI SDK → `gpt-5.4` (1M context, best general coding). Frontier
  multi-step research → `gpt-5.5` (~2× the cost of 5.4).
- Open-source / OpenCode agents → `fw-glm-5.1` (open-weights default). Longer
  context → `kimi-k2.6` (262K).
- High-volume batch (thousands of abstracts/papers) → `deepseek-v4-flash`
  (lowest cost, near GPT-4o quality).
- Image generation (figures, posters, diagrams) → `gpt-image-2`.
- Embeddings → `text-embedding-3-large`.

## Setup templates

### Claude Code (Anthropic-compatible endpoint, no /v1)

```bash
echo 'export CATION_AI_API_KEY="sk-PASTE-YOUR-KEY"' >> ~/.zshrc
echo 'export ANTHROPIC_BASE_URL="https://bmblx.bmi.osumc.edu/cation"' >> ~/.zshrc
# ANTHROPIC_AUTH_TOKEN (→ Authorization: Bearer) is the recommended auth for a proxy key.
echo 'export ANTHROPIC_AUTH_TOKEN="$CATION_AI_API_KEY"' >> ~/.zshrc
echo 'export ANTHROPIC_MODEL="claude-sonnet-4-6"' >> ~/.zshrc
# Pin the background/small-fast model to one that exists on the gateway, else Claude Code's
# haiku calls hit an unrouted default. (Current var name; ANTHROPIC_SMALL_FAST_MODEL is deprecated.)
echo 'export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-haiku-4-5"' >> ~/.zshrc
source ~/.zshrc
claude
```

### Codex (OpenAI-compatible endpoint, /v1) — merge into ~/.codex/config.toml

```toml
model_provider = "cation"
model = "gpt-5.4"
model_reasoning_effort = "medium"

[model_providers.cation]
name = "cation"
base_url = "https://bmblx.bmi.osumc.edu/cation/v1"
wire_api = "responses"
env_key = "CATION_AI_API_KEY"
```

### OpenCode — merge into ~/.config/opencode/opencode.json

```json
{
  "$schema": "https://opencode.ai/config.json",
  "model": "cation/fw-glm-5.1",
  "provider": {
    "cation": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "CATION Gateway",
      "options": {
        "name": "cation",
        "apiKey": "{env:CATION_AI_API_KEY}",
        "baseURL": "https://bmblx.bmi.osumc.edu/cation/v1"
      }
    }
  }
}
```

### Generic OpenAI SDK (Python)

```python
from openai import OpenAI
import os

client = OpenAI(
    base_url="https://bmblx.bmi.osumc.edu/cation/v1",
    api_key=os.environ["CATION_AI_API_KEY"],
)
resp = client.chat.completions.create(
    model="gpt-5.4",
    messages=[{"role": "user", "content": "ping"}],
)
print(resp.choices[0].message.content)
```

## Verify (one-line test)

```bash
curl -s https://bmblx.bmi.osumc.edu/cation/v1/models \
  -H "Authorization: Bearer $CATION_AI_API_KEY" | head -c 400
```

## Links

- Onboarding dashboard: https://bmblx.bmi.osumc.edu/onboard/
- Model catalog (JSON): https://bmblx.bmi.osumc.edu/onboard/data/models.json
- Contact / request a key: Cankun.Wang@osumc.edu