Emissary documentation

Emissary is a platform for LLM model routing, post-training and inference. Route each request to the model best suited to handle it, post-train models of your own on your data, and serve base and fine-tuned models on managed GPU infrastructure. Emissary replaces slow, expensive frontier-model calls with small, purpose-built classifiers that return the same decision in milliseconds for a fraction of the cost, with no prompt engineering and no full-size model in the loop.

What can I do with Emissary?

Model routing

Send each request to the cheapest model that can do the job, behind one OpenAI- and Anthropic-compatible gateway URL. Typical savings are 30–50%.

What is model routing? · Get started · Reduce LLM costs

Post-training

Adapt a base model to your task with supervised fine-tuning (SFT) or reinforcement learning (GRPO), then compare checkpoints and deploy the best one.

What is post-training? · Quickstart · Datasets

Inference

Serve base models and fine-tuned LoRA adapters on managed GPUs, with autoscaling from a single instance to a cluster of replicas.

What is LLM inference? · Deployments · Autoscaling

Quickstart

The fastest way to get started is one of our quick start guides. Each walks you through building your first model three ways: the public playground (no signup), the dashboard UI, and the API.

• Create Your First Judge: score any text against criteria you define (for example helpfulness or factual accuracy) and get back a calibrated probability instead of a vague verdict.
• Create Your First ModelRouter: automatically send each query to the model best suited to handle it, so you stop hard-coding one model for every request.

Want a router that also accounts for cache state, effort and output length? Start with model routing: get started. Want to train your own models? See the post-training quickstart.

How do the pieces fit together?

Routing, post-training and inference form a loop. Every routed request leaves a trajectory, so you learn which requests are easy and which no model can handle. You serve the open models you rely on from dedicated capacity, and you post-train models of your own on that traffic. The router sends work to them when they're the better call, which makes routing cheaper again. See what emerges from routing.

Reference

• API reference: REST endpoints for projects, datasets, training jobs, engines and deployments.
• Glossary: definitions of the terms used across these docs.
• llms.txt: a plain-text index of these docs for LLMs and AI assistants.