my-ai-toolsDocumentation index

Code taste with OpenRouter

code-taste uses the OpenAI Node SDK with an optional OPENAI_BASE_URL, so you can point analysis and embeddings at OpenRouter instead of OpenAI directly. No code changes are required—only environment variables.

Prerequisites

Quick setup

  1. Copy the repo template (already tuned for OpenRouter):

    cp .env.example .env

  2. Edit .env:

    OPENAI_BASE_URL=https://openrouter.ai/api/v1
OPENAI_API_KEY=sk-or-v1-your-key-here
OPENAI_MODEL=openrouter/free
OPENAI_EMBEDDING_MODEL=nvidia/llama-nemotron-embed-vl-1b-v2:free

  3. Load env and run (Bun):

    bun install
set -a && source .env && set +a

bun run code-taste analyze jellydn
bun run code-taste export --format markdown

    Optional: GITHUB_TOKEN increases GitHub API rate limits when listing repos and downloading files.

Environment variables

Variable Required Purpose
OPENAI_API_KEY Yes OpenRouter key (same variable name as OpenAI; SDK convention)
OPENAI_BASE_URL For OpenRouter Set to https://openrouter.ai/api/v1
OPENAI_MODEL No Chat model for preference inference (default in code: gpt-4o-mini)
OPENAI_EMBEDDING_MODEL No Embedding model for chunk selection (default in code: text-embedding-3-small)
GITHUB_TOKEN No GitHub REST API (recommended for user-wide repo discovery)

Implementation: lib/code-taste/profile.ts constructs the client with apiKey and baseURL from these variables.

Choosing models

Browse OpenRouter models or the curated Free models collection (rankings updated from real usage; capacity can change over time).

code-taste calls two APIs on the same base URL:

  1. Embeddings — ranks repository chunks before LLM analysis (OPENAI_EMBEDDING_MODEL).
  2. Chat completions — extracts preferences as JSON (OPENAI_MODEL, response_format: json_object in lib/code-taste/profile.ts).

Recommended free stack (matches .env.example)

Step Variable Suggested value Notes
API base OPENAI_BASE_URL https://openrouter.ai/api/v1 Required for OpenRouter
Key OPENAI_API_KEY sk-or-v1-... From OpenRouter keys
Embeddings OPENAI_EMBEDDING_MODEL nvidia/llama-nemotron-embed-vl-1b-v2:free Llama Nemotron Embed VL 1B V2 — text + image document embedding; works for text-only chunk strings
Analysis OPENAI_MODEL openrouter/free OpenRouter free router — picks an available free model per request

Example .env (copy-paste):

OPENAI_BASE_URL=https://openrouter.ai/api/v1
OPENAI_API_KEY=sk-or-v1-...
OPENAI_MODEL=openrouter/free
OPENAI_EMBEDDING_MODEL=nvidia/llama-nemotron-embed-vl-1b-v2:free

If openrouter/free or embeddings fail

Try a fixed free chat model from the free collection that supports tool calling / structured or JSON-style output, for example:

  • tencent/hy3:free — reasoning / agent workflows, 262K context
  • nvidia/nemotron-3-ultra:free — long context, strong on multi-step tasks
  • openai/gpt-oss-20b:free — structured outputs / function calling
  • Poolside Laguna variants (poolside/...:free) — agentic coding (check exact model id on OpenRouter)

Keep OPENAI_EMBEDDING_MODEL on an embedding-specific free model (e.g. nvidia/llama-nemotron-embed-vl-1b-v2:free). Do not point embeddings at chat-only models.

Free tier caveats

  • OpenRouter documents free inference as capacity they support but do not guarantee forever.
  • Some free providers (e.g. Poolside) may use inputs/outputs for training when used at $0 — read each model’s card before production use.
  • Free routers can return different backends per call; for reproducible taste profiles, pin a single OPENAI_MODEL id instead of openrouter/free.

The repository's .env.example supplies these OpenRouter settings. Export them before running code-taste analyze; Docker does not configure this CLI. Unlike the public-only repository chat indexer, code-taste can use GITHUB_TOKEN for authenticated GitHub API requests.

Review and retain a profile

Analysis writes .code-taste/profile.json and CODING_TASTE.md. The Markdown export omits evidence, so review the JSON evidence before treating inferred preferences as project rules. .code-taste/ and CODING_TASTE.json are ignored by Git; CODING_TASTE.md can be committed after review. These inferred profiles are separate from user-memory MCP, which stores only explicit user preferences.

Troubleshooting

Symptom What to check
OPENAI_API_KEY is required Export key after sourcing .env
401 / invalid key Key is OpenRouter sk-or-v1-..., not an OpenAI sk-... unless you use OpenAI’s URL
Embedding errors OPENAI_EMBEDDING_MODEL must exist on OpenRouter (or switch OPENAI_BASE_URL back to OpenAI for embeddings only—mixed providers need two runs or a single provider that offers both)
GitHub rate limit Set GITHUB_TOKEN with read access to the repositories you intend to analyze; do not grant write access for analysis

Related