open-notebook

We recommend all users upgrade. v1.15.0 brings the new "Quiet Green" visual identity, three new provider options (SiliconFlow, Z.ai, MiniMax text-to-speech), and the latest Esperanto, content-core and podcast-creator releases. Behind them is a long list of fixes to failures that used to be silent: podcasts that never appeared, sources stuck as "Processing...", workers that stalled, answers cut short. Validated through the full release process: backend and frontend suites, a fresh-install and an upgrade-from-1.14.0 test against the published images, and an end-to-end pass over the running app (API and UI), including podcasts on OpenAI, Anthropic, Gemini and DeepSeek.

🔒 Security

• Credential encryption now uses PBKDF2-HMAC-SHA256 (600k iterations) with versioned ciphertext. New keys are written in the new pbkdf2v1: format, and existing keys keep decrypting. A one-shot POST /api/credentials/migrate-encryption upgrades stored keys (idempotent, fail-closed per record). See the downgrade note below (#1317, #1360)
• Dependency security update: next 16.3.8 (critical RCE advisory in next/og; Open Notebook doesn't use next/og), axios, postcss, brace-expansion, and Python anyio, pyjwt, tornado, urllib3, virtualenv. Security floors are now pinned so a future lock can't resolve back to vulnerable versions (#1404, #1414)
• The pillow override is gone: podcast-creator no longer pulls in moviepy, and content-core now requires pillow >= 12.3.0 itself (#1440)

✨ New features

• "Quiet Green" visual identity. A new design foundation co-designed with the community in Discussion #1202 (a token system for color, type, radius and shadow; Bricolage Grotesque, Instrument Sans and Spline Sans Mono), applied screen by screen: app shell, notebook workspace, sources library, dialogs, models, settings, podcasts and search. Works in light and dark themes (#1202)
• New providers: SiliconFlow (hosted DeepSeek, Qwen, GLM and Kimi; mainland-China accounts set the credential's Base URL to https://api.siliconflow.cn/v1) and Z.ai (the GLM family), both with connection test and model discovery (#1409, #1437)
• MiniMax text-to-speech. MiniMax voices can now be used in podcast speaker profiles (speech-2.8-hd, speech-2.8-turbo), and MiniMax moves to MiniMax-M3 with a 1M-token context. Region-specific keys can set MINIMAX_BASE_URL / the credential Base URL (#1438)
• OpenDocument uploads: .odt, .ods (one table per sheet) and .odp (one block per slide), plus .htm files (#1439)
• YouTube when YouTube blocks your server: set CCORE_YOUTUBE_PROXY (a residential proxy) or CCORE_YOUTUBE_COOKIES_FILE (a browser cookies.txt) on the worker to get transcripts through. CCORE_AUDIO_SEGMENT_MINUTES controls how long audio is split before transcription (#1439)
• Podcasts in your language again: episode profiles' language setting is honored for outlines and transcripts (#1334)
• Podcasts sound more natural, with a short (~400 ms) pause between speaker turns, and audio combining no longer cuts off the end of a turn (#1446)
• Podcast failures now explain themselves: a wrong speaker name, a voice the TTS model doesn't support, or a truncated response each get a specific hint (#1334)
• The search page explains what text and vector search each cover; a "New Source" button sits on the empty sources state; the "API Keys" settings page is now "Models" (#1295, #1165, #1313)

🐛 Notable fixes

• Podcasts work on a fresh install. The seeded speaker profiles ship without a voice model, and any unconfigured profile used to fail every podcast with validation errors for SpeakerConfig, even one generated with a fully configured profile (#1450)
• Podcasts work with models that don't support JSON schema output (DeepSeek, older vLLM / llama.cpp servers). Outlines and transcripts use schema-validated output and fall back to plain JSON when an endpoint rejects it. Prompts no longer show

We recommend all users upgrade. v1.14.0 is a providers-and-plumbing release: six new ways to connect models, a DNS-rebinding hardening pass on every outbound provider call, and fixes to two failure modes that could silently break a working install after a redeploy. Validated through the full release process — backend/frontend suites, a fresh-install and an upgrade test against real Docker images, and an end-to-end pass over the running app.

🔒 Security hardening

• Outbound provider requests are now pinned to a DNS-vetted IP. Validating a URL and then handing it to the HTTP client left a rebinding window: the hostname was resolved again at connect time, so a name that passed validation could point at a cloud metadata endpoint on the actual connection. Resolution now happens once, dangerous addresses are rejected, and the request is pinned to the vetted address while Host and TLS SNI keep the original hostname — so certificate verification and routing are unaffected. Applied to credential saves, connection tests and model discovery (#1063)
• Private IPs, localhost, host.docker.internal, LAN hostnames and Tailscale addresses remain fully supported — this is a self-hosted application and local providers are a first-class use case. That guarantee is now pinned by regression tests rather than left implicit (#1193)
• Error text returned by the sources API is capped, so internal exception detail from extraction libraries no longer reaches clients (#1136)
• Resolved all open Dependabot alerts: mcp → 1.28.1 (3 high, transitive via content-core) and axios → 1.18.1 (#1191)

✨ New features

• Four new AI providers: Cohere (language + embedding via its native v2 API), PayPerQ / PPQ (multi-modality gateway — language, embedding, STT, TTS), Novita (OpenAI-compatible LLM gateway), and speech-to-text on the existing Deepgram provider (#1170)
• oMLX as a first-class provider — native Esperanto profile, default base URL http://localhost:11435/v1, connection test and model discovery, no remapping to openai-compatible (#1048)
• OpenRouter now does text-to-speech and speech-to-text, not just LLMs and embeddings. Note that its TTS model uses Microsoft neural voice names (e.g. en-US-AvaNeural), not OpenAI's alloy/nova set (#987)
• anthropic_compatible credential provider for endpoints that speak only the Anthropic Messages format, such as Kimi (#675)
• Docling formula & vision enrichment toggles in Settings → Content Processing — extract mathematical formulas as structured markup, and describe images / extract chart data with a vision model. Both opt-in and off by default; vision is significantly slower (#1131)
• Worker concurrency is configurable via OPEN_NOTEBOOK_WORKER_MAX_TASKS (default 5). Set it to 1 on single-GPU or local-LLM setups that were being overloaded by parallel source processing (#893)
• Newer ElevenLabs models in discovery — eleven_v3, eleven_flash_v2_5, eleven_flash_v2, and scribe_v2 for STT (#1167)
• Podcast episode profiles expose an optional "Max output tokens" field in the UI, with help text explaining it caps output per generation step (outline + transcript) and defaults to the built-in values when blank (#991)

⚡ Performance

• Typing in the chat composer no longer lags in long conversations. Composer state is isolated from the message list and each row is memoized, so a keystroke no longer re-renders the whole history. Most noticeable for IME users (Japanese, Chinese, Korean), where every composition event was triggering a full re-render (#1147)

🐛 Notable fixes

• A selected content-processing engine whose runtime isn't installed no longer breaks ingestion silently. The engine choice lives in the database; the opt-in runtimes that serve it come from environment flags. A redeploy that drops OPEN_NOTEBOOK_ENABLE_CRAWL4AI/_DOCLING therefore left the setting pointing at nothing, and every extraction failed with an u

We recommend all users upgrade. v1.13.0 is our biggest content-processing upgrade yet: it rebuilds extraction on top of Content Core 2.0 (a major architecture release), adds opt-in Docling (OCR + layout-aware parsing) and Crawl4AI (JavaScript rendering) engines, brings EPUB and Reddit support, moves PDF extraction to MIT-licensed pdfplumber, fixes audio extraction, and keeps the default image lean by making the heavy runtimes opt-in. Plus a batch of podcast correctness fixes. Validated with the release confidence process: full backend/frontend suites, the fresh-install + upgrade image gate, and manual verification of the opt-in runtimes, real podcast/TTS and PDF extraction on the pushed image.

✨ A major content-processing upgrade (Content Core 2.0)

Extraction now runs on content-core 2.x — a ground-up refactor of the extraction stack. What it means for you:

• Docling document engine (opt-in) — real OCR for scanned PDFs and images plus layout-aware parsing. Toggle OCR in Settings → Content Processing (on by default; turn it off to speed up text-native docs). (#1104)
• Crawl4AI URL engine (opt-in) — renders JavaScript-heavy pages locally with no API key, or offloads to a Crawl4AI server (CRAWL4AI_API_URL). Selectable alongside Firecrawl, Jina and Simple. And your persisted engine choice now actually takes effect — the source-processing graph reads the saved Content Settings and passes them to content-core (it previously always ran auto and silently ignored the selection). (#432)
• New formats — EPUB books are now supported, and Reddit links extract the post and its comments automatically.
• Better PDF & audio — PDF extraction moves to MIT-licensed pdfplumber (replacing AGPL PyMuPDF), and audio moves to direct ffmpeg calls — faster, and it fixes extraction from MP3 files carrying chapter metadata. (#1103)
• Faster, clearer failures — unsupported uploads are rejected instantly at ingestion with a 415 naming the detected type, instead of a background job retrying for ~an hour. (#975)

🪶 Lean by default, powerful on demand

• The heavy runtimes are opt-in. Docling and local Crawl4AI are no longer bundled into every image — they install automatically on first container startup when enabled. Set OPEN_NOTEBOOK_ENABLE_DOCLING=true and/or OPEN_NOTEBOOK_ENABLE_CRAWL4AI=true; downloads are cached on the /app/data volume (only the first boot is slow), and a failed install degrades gracefully (the app still starts, the engine is reported unavailable) with loud logs. A new GET /api/capabilities probe reports what's actually installed, and Settings → Content Processing disables the Docling/Crawl4AI engines and the OCR toggle (with an env-var hint) until their runtime is available — so the UI never advertises an engine that isn't there. A remote Crawl4AI server needs no local install. Recorded as ADR-007. (#1122)
• Source extraction now logs the effective engine (url_engine / document_engine / docling_ocr) at INFO right before it runs, so you can confirm which engine actually processed a source instead of guessing. (#1125)

🐛 Notable fixes

• Podcast correctness: renaming a speaker profile no longer breaks the episode profiles using it (references are stored by record ID, not name); episode audio paths are stored relative to the podcasts folder so previously generated episodes survive a DATA_FOLDER relocation; episode cards resolve and show the real outline/transcript/speaker model instead of "— / —"; and podcast generation now honors an explicitly supplied speaker_profile all the way into the generation command. (#630, #1030, #1114, #1058)
• Clicking a chat/Ask citation that points at a deleted source, insight or note now shows a friendly "this content no longer exists" state in all three dialogs. (#455)
• Source insights get created/updated timestamps and the API returns null (not the literal string `"N

We recommend all users upgrade. v1.12.0 is a maintainability and correctness release: a full-repo cleanup (20+ focused PRs) that removed ~3,000 lines of dead code, fixed real user-facing bugs along the way, and turned on the quality gates that keep them fixed — validated end-to-end by the release test process (full suites, API+UI smoke, fresh-install and upgrade tests on the real images).

✨ New features

• Live Anthropic model discovery — "Discover models" for Anthropic credentials now queries Anthropic's models API instead of a hardcoded 2024-era list, so new Claude models appear as soon as they ship (#1070)
• GET /api/providers — new endpoint exposing provider metadata (modalities, docs, env-configured status) from a single backend registry (#1075)
• Math rendering in chat — prompts now steer models to emit $$...$$/$...$ so formulas render via KaTeX instead of raw LaTeX code blocks (#1051)

🐛 Notable fixes

• Errors now tell you what's wrong: a missing/misconfigured model returns a clear 422 with guidance (and not-found resources a 404) instead of a generic "Internal Server Error" — across all API routes (#1078)
• Translations interpolate correctly in all 14 languages: counters, names and plurals no longer risk showing literal {count}/{name} placeholders; eight podcast toasts that always showed {name} are fixed (#1074, #1077)
• Podcast dialog: rapid source toggling no longer causes a request storm with flickering token counts, and the dialog closes deterministically after the episode list refreshes (#1067)

🧹 Code health (why this release exists)

• ~3,000 lines of dead code removed, including the entire leftover Streamlit-era service layer (#1053, #1054, #1056)
• CI now gates every PR on ruff, ESLint, the frontend production build and mypy at zero errors (was: 197 baseline errors, not enforced) (#1062, #1068, #1076)
• One provider registry drives env config, modalities, connection tests and discovery — adding a provider went from ~6 hand-synced locations to 2, both test-enforced (#1075)
• Heavy copy-paste deduplicated across the chat routers, sources router, embedding commands and model discovery (#1069–#1072)
• Both Docker images are now built from a single multi-stage Dockerfile, eliminating drift between the regular and single-container variants — published image names and tags are unchanged (#1066)
• Dependency diet: unused langchain-community and langchain-deepseek removed, upper bounds on the langchain family, and an undeclared transitive dependency made explicit (#1073)

⚠️ Behavior changes for self-hosters

1. API error codes are now accurate — if you script against the API and matched on 500 for configuration mistakes, update your handling: configuration errors return 422, missing resources 404, provider rate limits 429, upstream provider failures 502 (#1078).
2. POST /api/notebooks/{id}/context was removed — it duplicated POST /api/chat/context and had no known callers. If you used it, switch to POST /api/chat/context with body {notebook_id, context_config} (#1079).
3. Upgrading from a version older than 1.6 with embedding jobs still queued: drain the queue on a 1.x release before upgrading past this version — the pre-1.6 job compatibility shims were removed (#1056).

🙏 Thanks

This cycle was a focused maintenance sprint by the core team, but it stands on the community's shoulders: thank you to everyone who reported the issues that shaped the cleanup backlog and the release checklist — your reports are what turned a "refactor" into a list of real bugs to fix. Special mention to the contributors whose earlier security work (#1042, #1046, #1049) set the quality bar this release enforces in CI.

Full details in the CHANGELOG.