Cogfer

A library for calling LLMs through one typed request model, stream contract, and error taxonomy.

Supports:

• OpenAI (Responses and Chat Completions)
• ChatGPT subscriptions
• SpaceXAI
• SpaceXAI subscriptions
• OpenRouter
• Anthropic
• Anthropic on Amazon Bedrock
• OpenAI on Amazon Bedrock
• Gemini

Bedrock support covers the Anthropic and OpenAI dialects; other Bedrock vendors (which would need the Converse API) are out of scope for now.

Quick start

[dependencies]
cogfer = { git = "https://github.com/caido/cogfer" }

use cogfer::{Client, Credentials, Message, ProviderConfig, Request};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Once per process, unless the host already installed a Rustls provider.
    cogfer::transport::install_default_crypto_provider();

    let client = Client::builder().build()?;
    let provider = client.provider(ProviderConfig::openai_responses(
        Credentials::api_key(std::env::var("OPENAI_API_KEY")?),
    ))?;
    let model = provider.language_model("gpt-5.6-luna");

    let request = Request::builder()
        .message(Message::user("Why is the sky blue?"))
        .build();
    println!("{}", model.generate(request).await?.text());
    Ok(())
}

Streaming:

use futures_util::StreamExt;

let mut stream = model.stream(request).await?;
while let Some(event) = stream.next().await {
    if let cogfer::StreamEvent::TextDelta { delta, .. } = event {
        print!("{delta}");
    }
}

Events are designed for agentic environments and map cleanly onto AI SDK stream parts. A typical event sequence:

StreamStart { warnings: [] }
ResponseMetadata { id: "resp_abc123", model: "gpt-5.6-luna", .. }
ReasoningStart { id: "rs_0" }
ReasoningDelta { id: "rs_0", delta: "The user asks why the sky is blue" }
ReasoningDelta { id: "rs_0", delta: " — Rayleigh scattering explains it." }
ReasoningEnd { id: "rs_0", part: ReasoningPart { .. } }
TextStart { id: "0", .. }
TextDelta { id: "0", delta: "The sky is blue" }
TextDelta { id: "0", delta: " because sunlight scatters" }
TextDelta { id: "0", delta: " off air molecules." }
TextEnd { id: "0", .. }
Finish { reason: Stop, usage: Usage { input_tokens: 13, output_tokens: 12, .. }, .. }

Features

• Blocking and streaming completions
• Tool calls and reasoning, replayable across turns
• Structured output, context compaction
• Normalized errors, usage, and warnings
• ChatGPT and SpaceXAI subscription OAuth
• Pluggable HTTP transport

Examples and tests

OPENAI_API_KEY=... cargo run --example stream
cargo test   # no network tests
OPENAI_API_KEY=... cargo test --test lib live::openai -- --ignored --nocapture # will use OPENAI_API_KEY and call openai for tests