The Raven system: routing, reasoning, and seeing

A model is easy to demo and hard to run. The difference is the system around it. This post describes how the Raven system turns a request into an answer: what gets routed, what gets inspected, and what never gets stored.

One interface, many paths

Every request to Raven passes through a routing layer before any model sees it. The layer reads the request the way a mail sorter reads envelopes: text flows one path, images another, and a request that mixes both is split so each part travels the path built for it. Callers never manage this. A request that carries an image gets vision; a plain question gets the text path. The caller picked a model; the system picks the road.

Reasoning you can inspect

The Raven models stream a reasoning channel alongside their answers. In DCode and the Raven app it renders as a thinking block: collapsed by default, escaped plain text, one click to expand. You can watch the plan form before the answer lands, which changes how you trust it.

The reasoning stream is also a safety surface. On turns that look like attempts to extract the model's instructions or identity, the system suppresses the thinking channel entirely and delivers only the answer. The visible reply stays normal; the internals stay private.

Tools, in a loop

Agentic work is a loop: plan, call a tool, read the result, revise, continue. Raven models call tools natively, and the system keeps the loop honest. Tool-call arguments stream through the same inspection and safety layers as prose, so a long agent session is as controllable as a single question. In DCode the loop reads files, edits them, runs your commands, and reports back with receipts.

What we keep, and what we never keep

The system meters what it must to operate: request counts, token totals, model tier, timestamps. It does not log or store the content of your prompts or completions. Rate windows, credit balances, and plan entitlements are enforced per request, per key, per user. The result is a service a privacy reviewer can read end to end.

What comes next

The routing layer is where our research investment goes: smarter selection between Flash and Max, richer enrichment for raw API calls, and measurement that tells us when a route made an answer better and when it only made it cheaper. The models will improve. The system around them is built to improve with them.