Bring your own keys, keep every conversation on your machine. Anthropic, OpenAI, Gemini, xAI, Mistral, Groq, DeepSeek, OpenRouter, Ollama, or any OpenAI-compatible endpoint — side by side.
v0.7.7
Decision models, in Studio. A new Decide tab asks a model typed questions about anything you paste in — yes / no, pick one, or score on a scale — and shows a probability for every answer instead of a paragraph of text. Works with TypeSafe Jev, Cloudflare Clef and Clef-flash on Workers AI, and OpenAI's Decisions API on your existing OpenAI key.
Bring your own decision endpoint. Any server that speaks the System One API can be added with a base URL and a key under Settings → Connections → Decision.
v0.7.6
Support for additional fields in pricing models. Support for maxIn, maxOut, and notes fields has been added to pricing models. The UI has been enhanced with tooltips to provide additional context for these fields.
Sync pricing catalog. Synced the pricing catalog using the public price list.
Bug fixes.
v0.7.5
A generic LLM abstraction is now the main layer. The app now uses a generic abstraction for agents, embeddings, and vector indexing instead of provider-specific implementations.
Assistant workflows were refactored. Tool approvals are now thread-safe, and agent runtime/memory/tool hooks were tightened up.
v0.7.4
Images and speech share one Studio. Image generation, text-to-speech, and transcription now live behind a single sidebar entry. One header, one switcher — Image, Text → Speech, Speech → Text — instead of two rail items and a tab hidden inside one of them. Your gallery, clips, and transcripts are untouched.
Skills moved into the Library. The prompt library is now Library, with a Skills tab next to Prompts. Skills get a full page instead of a popup buried in Settings → Tools: install from a git URL, open the folder, toggle each one, remove several at once.
v0.7.3
Paste anything, straight into chat. Copy an image, PDF, or text file anywhere on your machine and hit paste. Images attach as images; PDFs and text files are read into your prompt as context. No attach button hunting.
Show, don't just tell, your image model. Attach a reference image in Images and describe the change. Image models take text only, so a vision model reads your reference and folds the description into the prompt behind the scenes — your prompt box stays as you typed it. Choose the reader under Tools → Reference image reader. Billed as one chat turn.
Generated images now show up in your bill. Priced per image by model, size, and quality — GPT Image 2 / 1.5 / 1 / 1 Mini, DALL·E 2 and 3, Imagen, Stability — across totals, daily chart, per-model breakdown, and CSV export, plus a new Images counter. Past images counted retroactively; per-image cost shows in the gallery. Costs always recompute from the current price table, so a price fix repairs your history.
v0.7.0
A fresh look. New color scheme across the app and the icons.
Perplexity search. Add your key under Keys & security and pick it as the search backend.
Nested lists that actually nest. Multi-level bullets and numbers keep their structure, with distinct markers per level (hollow bullets, letters, roman numerals). Mixed nesting works; wrapped lines stay in their list item.
Master-password changes can't strand you. Turning protection on, changing it, or turning it off is now all-or-nothing. Migrations wait for the database file lock instead of failing instantly, and the app only records "password required" once the database is genuinely encrypted with it. A failed attempt retries cleanly on next launch.
Bug fixes.
v0.6.x
Agent Skills. Teach the model your workflows using the open Agent Skills format — a folder with SKILL.md, reference files, and scripts. The model loads full instructions on demand mid-chat and can run bundled scripts, each with your approval (60s timeout, working directory pinned to the skill folder). Works with any provider that supports tool calls.
Bring your existing skills. Drop folders into the skills directory, clone a repo from a git URL, or reuse what you have — ~/.claude/skills and ~/.agents/skills import automatically as read-only.
Skills manager. Tools page shows installed / enabled / invalid at a glance; Manage skills handles the rest — toggle per skill, add from URL, open the folder, bulk-remove. New skills start disabled. Invalid ones tell you why instead of vanishing.
Master password. Your password _is_ the database key, via Argon2id — nothing auto-unlocks, and forgetting it means the data stays unreadable. Set, change, or remove it under Keys & security. A lock screen gates launch, with throttled retries and a last-resort "reset app" that wipes local data.
Prompt firewall. Strip sensitive data before it leaves your device, restore it in the reply. Detection is fully on-device: emails, phone numbers, credit cards (Luhn-checked), IBANs, SSNs, IPs, and URLs become stable placeholders ([EMAIL_1]); API keys block the send outright. Add a custom dictionary for codenames and customer names, plus an allowlist. Consistent placeholders keep the model reasoning about relationships; restore is echo-tolerant and streaming-safe. Off by default — enable globally or per chat.
On-device ML detection. An optional local NER model catches what patterns miss — people, organizations, locations. Runs on your CPU, adds some latency per message.
Firewall tester. See the full round trip on sample text: your input, exactly what gets sent, and the restored result, with detected entity types and the placeholder map. Nothing leaves your machine.
v0.5.x
Encrypted at rest (v0.5.3). Your local database is fully encrypted on disk with SQLCipher (AES-256). The key is generated on your machine and stored in the OS keychain, so the app still opens instantly with no password. Existing databases migrate on first launch; backups stay encrypted.
Code formatting (v0.5.2). Codebase cleanup and consistent formatting.