Changelog

Every release, on the gauge.

Penweir is named after a weir, so we read our versions the way the Netherlands reads water: as a level on a gauge. Scroll down and the water follows the release you are reading.

How to read a version

Metres, decimetres, centimetres.

Each number in a version is a mark on the staff. 1.4.0 stands at +1.40.

  1. Major · metreA new generation of Penweir.
  2. Minor · decimetreNew capabilities.
  3. Patch · centimetreImprovements and fixes.

Level +1.40MinorCurrent

Better at finding what people really write

A detection release. Penweir finds IBANs, BSNs and payment cards as people write them, no longer mistakes the password in a URL for an email address, and lets custom detectors redact one part of a match. It follows the research on this site.

This is the version that runs for companies with early access today, until Penweir's commercial release.

Redaction

  • IBANs with dashes, no-break spaces or a line break between the groups are found, and the checksum validators ignore those separators. On a test corpus of 36,000 synthetic documents, recall on messy IBANs went from 0.45 to 0.99.
  • BSNs written 3-3-3 are found next to the usual 4-2-3, and payment cards with a line break between groups too (0.77 to 0.97 on messy text).
  • The password in a URL with credentials is no longer also reported as an email address. Email precision went from 0.50 to 1.00.
  • A known secret format, such as a JWT after API_TOKEN=, is one detection and not several fragments.

Configuration

  • Custom detectors accept an optional group: only that part of the match is redacted and checked. This finds a KvK number by its label without the label becoming part of the match.

Gateway

  • Requests refused at the concurrency cap are counted and logged, and new load tests check fast 503 refusals with Retry-After and one audit event for every admitted request.

Dashboard

  • The release notes moved out of the application to this page, and the System page links here.

Level +1.30MinorReleased 15 November 2026

Demo mode, and a dashboard that comes alive

The first release with a Demo mode: without a license Penweir runs with fixed limits and no time limit, and a license key lifts them. It also adds the Insights and Security pages, the live dashboard and a rebuilt Settings page.

Demo mode and licenses

  • Without a license: 1 provider, 2 applications, 1 pipeline, 3 detector steps and 50 requests a day. Everything else works in full.
  • A license is a PNW1- key bound to the installation's Instance ID and checked offline. Paste it in Settings and the limits lift at once, without a restart.

The live dashboard

  • A new Overview: requests, values held back, the time Penweir adds, and payload stored: 0 B.
  • The audit log updates while you watch, and the pipeline flow replays real requests from their counts, never from content.
  • Light and dark themes, a command palette and an accessibility pass.

Insights

  • Provider health, detection trends and a cost estimate per application and model.
  • Key rotation and hygiene, pipeline templates and a content-free diagnostics bundle.
  • An optional alerts webhook that sends counts and identifiers only.

Security page

  • Every security control in one place, mapped to the OWASP Top 10 for LLM Applications.
  • AI controls such as prompt-injection signatures, invisible Unicode and a system-prompt canary, each off, monitor or enforce.
  • Weakening a critical control asks for its name and your password.

Server and release

  • PENWEIR_WORKERS runs up to 16 gateway processes.
  • Clear startup errors instead of a traceback, and stricter file permissions for the database.
  • A reproducible Demo bundle, image tarballs and checksums with every release.

Level +1.15Patch

Faster, and steadier under load

A performance release. Penweir adds less time to each request, serves 1.5 to 2 times as many requests per process, stays responsive during very large requests, and streams answers without a visible delay.

Gateway

  • Added latency for a 4 KB prompt went from 3.2 to 1.3 ms at the median, and from 4.9 to 1.9 ms at p95.
  • Hosts are resolved when a connection opens, so a slow DNS server no longer slows every request.
  • Large requests are scanned in a worker thread, so other requests and health checks keep being served.

Streaming

  • Response steps hold back only text that could still become a match, so the first words of an answer arrive about 60 tokens sooner.

Security and fixes

  • The SSRF check now connects to the address it checked.
  • A KvK number streamed after its keyword is now redacted by a response step.

Level +1.10Minor

Hardened after an audit

A security release, following an audit of 1.0.0. It closes gaps in what the redaction steps scan, hardens admin login and secrets, and adds rate limits. Upgrading needs no migration.

Redaction

  • Tool descriptions and parameters, response schemas, stop sequences, message names and tool-call ids are now scanned.
  • Response steps also scan tool-call arguments and refusals.
  • The email detector runs in linear time.

Gateway

  • Rate limits per key (600 a minute by default) and for the whole gateway (6,000).
  • NAT64, 6to4 and Teredo addresses are judged by the IPv4 address they reach.
  • Connector answers are read up to a size limit.

Admin and secrets

  • A stored secret is bound to the first host it is sent to.
  • Sensitive actions need your password from the last ten minutes.
  • The login lockout holds under parallel guesses, and a used TOTP code is refused.

Deployment

  • The Helm chart mounts the master key as a read-only file.
  • With the network policy on, the admin port has its own closed rule.

Level +1.00First release

The first release

A self-hosted AI gateway between your systems and LLM providers. It decides per policy what passes, and logs metadata only.

Gateway

  • OpenAI-compatible endpoints with streaming and tool calling. The official SDK works with only base_url changed.
  • OpenAI-compatible and Anthropic providers, with an allow-list, a preferred provider and fallback.
  • Applications with API keys that are shown once, stored hashed and revocable.

Redaction

  • Detectors for email, phone, IBAN, BSN, KvK, VAT, postcodes, payment cards, IP addresses, credentials in URLs and secrets, with checksums where they exist.
  • Redact, mask, tokenize, hash or block. Tokens are restored in the answer, also while streaming.

Connectors

  • Read-only Odoo, SQL and mailbox connectors, run by the gateway as tools. Every result is filtered before the model sees it.

Security and privacy

  • No payload storage, enforced by a canary test. The audit log holds metadata only. No telemetry.
  • Secrets encrypted at rest, argon2id passwords, optional two-factor login and a strict CSP.
  • A non-root container with a read-only filesystem and no shell, and an SBOM with every release.

Deployment

  • One image for amd64 and arm64, Docker Compose and a Helm chart.