Onwards

A Rust-based AI Gateway that provides a unified interface for routing requests to OpenAI-compatible targets. The goal is to be as "transparent" as possible.

Read the full documentation

Quickstart

Create a config.json:

{
  "targets": {
    "gpt-4": {
      "url": "https://api.openai.com",
      "onwards_key": "sk-your-openai-key",
      "onwards_model": "gpt-4"
    }
  }
}

Start the gateway:

cargo run -- -f config.json

Send a request:

curl -X POST http://localhost:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Features

Unified routing to any OpenAI-compatible provider
Hot-reloading configuration with automatic file watching
Authentication with global and per-target API keys
Rate limiting and concurrency limiting (per-target and per-key)
Load balancing with weighted random and priority strategies
Automatic failover across multiple providers
Strict mode for request validation and error standardization
Response sanitization for OpenAI schema compliance
Prometheus metrics
Custom response headers
Optional multi-step Open Responses orchestration loop (multi-step feature)

Multi-Step Open Responses

The multi-step Cargo feature adds an orchestration loop that drives /v1/responses-style requests through one or more upstream model calls, optional server-side tool calls, and back — without onwards itself owning any persistence. The loop is a free function (run_response_loop) over two traits a consumer plugs in:

MultiStepStore — records steps, walks the chain, runs the transition function (next_action_for: given the chain so far, emit the next step, complete, or fail), and assembles the final response. Implementing this trait is the host application's job; the reference implementation in Doubleword's control-layer is backed by fusillade for durable per-step rows.
ToolExecutor — the same trait onwards already uses for single-step tool injection. The loop calls execute for ToolKind::Http tools; ToolKind::Agent tools cause the loop to recurse into a sub-loop under that tool's step id.

Model calls fire through fusillade::HttpClient so each per-step HTTP request carries the X-Fusillade-Request-Id header that downstream analytics keys on. Streaming responses propagate live token deltas to an optional EventSink via fusillade's StreamEventCallback hook, so warm-path SSE clients still see chunks arrive in real time while fusillade owns reassembly into the canonical final body.

Enable

[dependencies]
onwards = { version = "0.27", features = ["multi-step"] }
fusillade = "17.0.2"

The feature is off by default — the simple /ai/v1/* proxy path doesn't need it and stays fusillade-free. Turning it on pulls in fusillade for RequestData, HttpClient, and the streaming-callback types.

Wire-up sketch

use onwards::{run_response_loop, LoopConfig, UpstreamTarget};
use std::sync::Arc;

let http_client = Arc::new(fusillade::ReqwestHttpClient::new(
    /* first_chunk_timeout */ std::time::Duration::from_secs(30),
    /* chunk_timeout */       std::time::Duration::from_secs(30),
    /* body_timeout */        std::time::Duration::from_secs(120),
    /* streamable_endpoints*/ vec!["/v1/chat/completions".into()],
));

let final_response = run_response_loop(
    &store,           // your MultiStepStore impl
    &tool_executor,   // your ToolExecutor impl
    &tool_ctx,        // RequestContext carrying resolved tool config
    &UpstreamTarget {
        endpoint: "https://api.openai.com".into(),
        path:     "/v1/chat/completions".into(),
        api_key:  Some(api_key.to_string()),
    },
    http_client,
    /* event_sink */  None, // Some(&sink) for streaming
    request_id,
    /* scope_parent */ None,
    LoopConfig::default(),
    /* depth */ 0,
).await?;

Each model_call step's RecordedStep carries a sub_request_id — the fusillade.requests row id the store created for that step — which the loop stamps on the outgoing HTTP fire so per-step analytics line up with the right row in the database.

Performance Tuning

Connection Pooling

Onwards uses HTTP connection pooling to dramatically improve performance under load by reusing connections instead of creating new ones for each request. This eliminates the 1:1 request-to-file-descriptor ratio and prevents TIME_WAIT connection accumulation.

Configure via config file:

{
  "targets": {
    "gpt-4": {
      "url": "https://api.openai.com",
      "onwards_key": "sk-your-openai-key"
    }
  },
  "http_pool": {
    "max_idle_per_host": 100,
    "idle_timeout_secs": 90
  }
}

When to increase pool-max-idle-per-host:

The pool limit is applied per upstream host. Choose based on your deployment pattern:

Scenario 1: Fan-out (Multiple Upstreams)

Example: Main server routes to 10+ different model providers

Recommendation: max_idle_per_host: 200-300
Why: Traffic spreads across many upstreams, each gets moderate volume
Math: 10 providers × 200 connections = 2,000 total pooled connections

# Fan-out configuration
{
  "http_pool": {
    "max_idle_per_host": 300,
    "idle_timeout_secs": 90
  }
}

Scenario 2: Single Upstream (High Concurrency)

Example: Gateway in front of a single vLLM server handling all traffic

Recommendation: max_idle_per_host: 1000-2000
Why: ALL traffic goes to one host - needs high capacity to avoid creating new connections
Math: Peak 2000 concurrent requests → 2000 pooled connections reused across all requests

# Single upstream configuration
{
  "http_pool": {
    "max_idle_per_host": 2000,
    "idle_timeout_secs": 120
  }
}

Default values: If http_pool is omitted, defaults are 100 max idle connections per host and 90 second timeout.

Rule of thumb: Set pool-max-idle-per-host >= your expected peak concurrent requests per upstream host. If the pool is too small, new connections will be created beyond the pool limit, reducing the performance benefit.

Idle Timeout

Why 90 seconds is optimal:

The pool-idle-timeout-secs setting (default: 90s) controls how long idle connections stay in the pool:

HTTP/2 standard: Recommends keeping connections alive for 2+ minutes
Cloud load balancers: Typically timeout after 60-120 seconds
90s balances:
- ✅ Long enough to reuse connections between request bursts
- ✅ Short enough to avoid holding stale connections upstream LBs have closed
- ✅ Prevents connection leak from forgotten idle connections

When to adjust:

Increase to 120s for more aggressive connection reuse (bursty traffic with gaps)
Decrease to 60s if you see connection errors (upstream closing connections sooner)

Monitoring Connection Usage

# Monitor file descriptor usage (Linux/macOS)
lsof -p $(pgrep onwards) | wc -l

# Check connection states
ss -s  # Linux
netstat -an | grep TIME_WAIT | wc -l  # macOS/Linux

Expected improvements with connection pooling:

Before: 1000+ file descriptors for 200 concurrent requests (1:1 ratio)
After: 150-300 file descriptors for 200 concurrent requests (10:1 reuse ratio)
TIME_WAIT connections drop from thousands to near-zero

Documentation

Full documentation is available at doublewordai.github.io/onwards, covering:

onwards 0.29.1