llm.api

Minimal-dependency LLM chat interface. Part of cornyverse.

Exports

Function Purpose
chat(prompt, model) Chat with any LLM
chat_openai(prompt) OpenAI GPT models
chat_openai_codex(prompt) OpenAI Codex via ChatGPT subscription auth
chat_claude(prompt) Anthropic Claude (API key)
chat_claude_oauth(prompt) Claude on a Claude subscription (OAuth)
chat_ollama(prompt) Local Ollama server
list_ollama_models() List Ollama models
llm_base(url) Set API endpoint
llm_key(key) Set API key

Providers

Usage

# Auto-detect provider from model
chat("Hello", model = "gpt-5.4-mini")
chat("Hello", model = "claude-3-5-sonnet-latest")
chat("Hello", model = "kimi-k2.5")

# Use convenience wrappers
chat_ollama("What is R?")
chat_claude("Explain machine learning")

# Explicit Moonshot/Kimi provider
chat("Write a fast parser in R", provider = "moonshot", model = "kimi-k2.5")

# ChatGPT subscription-backed Codex provider (log in once; see below)
chat_openai_codex("Write a small R function")
chat("Refactor this loop", provider = "openai_codex", model = "gpt-5.5")

# Claude subscription-backed provider (log in once; see below)
chat_claude_oauth("Write a haiku about R")
chat("Refactor this loop", provider = "anthropic_claude")

# Any OpenAI-compatible endpoint via a custom base URL (see below)
llm_base("https://openrouter.ai/api/v1")
chat("Refactor this loop", provider = "openai_compatible",
     model = "meta-llama/llama-3-70b-instruct")

# Provider-native web search (the model searches on its own when useful)
chat("What changed in the latest R release?", web_search = TRUE)

# Conversation history
result <- chat("Hi, I'm Troy")
chat("What's my name?", history = result$history)

# Streaming
chat("Write a story", stream = TRUE)

Set MOONSHOT_API_KEY to use Moonshot/Kimi without overriding your OpenAI credentials.

OpenAI Codex (ChatGPT subscription)

The openai_codex provider talks to Codex using your ChatGPT subscription instead of an API key. Authentication is a one-time device login; the token is cached and refreshed by tinyoauth, so you log in once and it persists across R sessions.

# One-time: device-code login. Prints a URL + code to authorize in a
# browser. The token is cached under tools::R_user_dir("tinyoauth").
openai_codex_login()

# Thereafter, just use the provider; credentials come from the cache
# and refresh automatically.
chat_openai_codex("Write a small R function")
chat("Refactor this", provider = "openai_codex", model = "gpt-5.5")

Models: gpt-5.5 (default), gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark.

To use an externally-obtained token instead of logging in, set OPENAI_CODEX_ACCESS_TOKEN (and optionally OPENAI_CODEX_ACCOUNT_ID); these override the cache.

Claude (Claude subscription)

The anthropic_claude provider drives Claude on a Claude Pro/Max subscription instead of an API key, the same OAuth pattern as openai_codex. The token is cached and refreshed by tinyoauth, so you log in once and it persists across R sessions.

# One-time: prints an authorization URL. Approve it and paste the code back.
claude_oauth_login()

# Thereafter, just use the provider; credentials come from the cache
# and refresh automatically.
chat_claude_oauth("Write a haiku about R")
chat("Refactor this", provider = "anthropic_claude")

To use an externally-obtained token instead of logging in, set ANTHROPIC_CLAUDE_ACCESS_TOKEN; it overrides the cache.

OpenAI-compatible endpoints

The openai_compatible provider targets any endpoint that speaks the OpenAI chat-completions format: OpenRouter, DeepSeek, DeepInfra, a local proxy, or a corporate gateway. Set the endpoint with llm_base() or the OPENAI_COMPATIBLE_BASE_URL environment variable; /chat/completions is appended, so include whatever /v1 prefix the gateway expects. The model id is passed through untouched and is required (there is no default).

llm_base("https://openrouter.ai/api/v1")
chat("Write a haiku about R",
     provider = "openai_compatible",
     model = "meta-llama/llama-3-70b-instruct")

The API key comes from OPENAI_COMPATIBLE_API_KEY, falling back to OPENAI_API_KEY; a keyless internal gateway (no Authorization header) works with neither set. A missing base URL or model fails fast with setup instructions rather than a curl error.

chat() and agent() take a web_search argument (FALSE by default, TRUE, or a list of allowed_domains / user_location options) that runs the model’s own server-side search rather than a separate search API. The model decides when to search; the result carries citations and searches.

It’s wired for the hosted providers: openai / openai_codex (the Responses web_search tool), anthropic / anthropic_claude (the Messages web_search tool), and moonshot (its $web_search builtin). Other providers ignore it with a warning.

res <- chat("What changed in the latest R release?", web_search = TRUE)
res$citations

Dependencies

curl, jsonlite, and tinyoauth (for Codex device login and token caching). No tidyverse, no compiled code.