Ragen: what we shipped in the first half of September

Ragen gets a new version on every merge to main, so the number on its own says very little. Underneath, September had two clear threads: an answer that shows where it came from, and a new interface.

An answer you can check

The largest block of changes was about citations. The model now answers with numbered footnotes, and the number is verified, so [3] cannot point at a source that was never in the context. A footnote in the answer is a clickable chip, and a sources block sits below the answer. The block also shows the page a fragment came from, because a chunk now remembers its page number.

Along with that came visibility into the process itself. The stream says what retrieval did, and the interface shows what the answer was based on. The reranker score travels all the way to the answer. A thread reopened later still cites its sources.

Scope selection was added. The message field lets you set how much a thread is allowed to search. Level 1 searches what is visible in the file list anyway.

While we were there, we fixed the counting. The worker counts the pages the parser found, not characters. The interface now says outright that page totals are approximate.

⌘K, or one field instead of four screens

The command palette grew over three days. First the ⌘K shortcut itself, along with a hint that it exists. Then document search, then actions and questions — the palette started not only finding things but running them. Finally came the reasoning: each result shows why it matches the query.

Settings got a single navigation rail and two sections.

New layout: design system v2

Underneath, a v2 token layer was built, in which one token has one meaning. Accessibility was improved. On a phone, the chat field and the panel got new room to breathe.

Who sees what

Registration stays closed until an administrator opens it. The registration toggle saves its own change.

Write limits set per organisation were added, and controls a restricted organisation would not get to use are hidden. Someone outside the organisation gets their own visibility scope.

Envelope encryption got a home of its own, the @ragenai/crypto package (v1.136.0). The key provider reads a documented master key and supports AWS KMS. S3 credentials stopped using the AWS_* prefix, which frees it up for actual Bedrock and KMS.

Ragen as an MCP server

On 5 September came apps/mcp, an MCP server exposing Ragen’s chat to other agents, together with the ragen_list_assistants tool (v1.123.0). We also added OpenTelemetry and a structured logger, with the caveat that the raw OTLP endpoint never reaches the logs, because it can carry a credential (v1.124.0). The API also stopped rejecting standard OpenAI chat-completion parameters (v1.124.0).

Analytics counts what the model actually saw

Knowledge analytics counts the documents an answer cited, not the ones that were retrieved (v1.135.0). We also record what the model was handed, not only what it ended up citing. The whole screen shares one time window, ratings and a staleness notice were added, and the API dropped out of the numbers — the screen says so outright.

create-ragen-app

The week closed with a scaffolder: create-ragen-app stands up a self-hosted instance with one command (v1.161.0).

We laid the groundwork a few days earlier. A typed environment contract shared across the apps (ADR-37, v1.125.0), a Prisma client generated from a shared schema (ADR-40 step 1, v1.131.0), apps/api reading .env.local from the repository root like the rest of the apps.

The changelog page this post comes from reads releases that already existed anyway (v1.157.0).

In brief

• Next 16.3.4 clears the critical advisory the audit kept tripping over (v1.136.1)
• The Playwright cache key never changed, so an old browser stayed forever (v1.139.2)
• cancelEmbedding is a real signal, embeddingState a query, and permanent workflow errors are marked non-retryable (v1.121.0)

Full list of releases: docs.ragen.ai/changelog

Related posts:

• From a question to an answer with a citation. You don’t have to take AI at its word.
• MCP — USB-C for AI. One assistant instead of five chatbots
• Cloud or on-premise? How to pick an AI deployment model you won’t regret

Want to see this on your own documents? Book a free demo — in 30 minutes we’ll show you Ragen on your scenario.