Note: this page pulls live from GitHub Releases. If it has not loaded yet, the entries below are a recent snapshot.
v0.11.18July 15, 2026Latest
New
Consent-first security overhaul, rolled up across ten releases: every approval is bound to the exact action it approves, auto-approving modes are an explicit opt-in taken at launch, a project you open can tighten your security posture but never loosen it, secrets are redacted from the tool timeline, and the app asks before exposing anything on your network, all independently cross-audited
Voice dictation: speak straight into the chat composer on the signed Mac build, with on-demand mic access, speech-to-text through a provider you already connected, and bundled local Whisper as the always-ready offline floor. Local voice models, Whisper and Kokoro, download with live progress
Task-completion notifications with an optional sound, quiet hours on by default, and per-chat focus control
Native agent terminal: every chat can open a real TUI over a PTY, so you can watch your agent work at the metal or take the keyboard yourself
Ignition, a new built-in assistant for building income-generating assets with a curated six-domain expert skill set, and Hermes profiles as selectable preset assistants
Portable exports: any assistant exports as a clean SKILL.md and any workflow as a portable package, both with credentials redacted automatically
Added
Custom model IDs per provider, a one-click model switch when a chat hits a context ceiling, and an accurate context meter sized from the real model window
Live agent activity in the sidebar with a live tree of sub-agents as they spin up and finish
Concierge files a detailed one-click GitHub report with screenshot and diagnostics, and a new /doctor command checks your workspace and config on demand
Core extensions management in Settings, an updater that matches your theme, and an About panel that checks for updates on open
Improved
True Continue: a stopped or capped run resumes the live turn instead of restarting, and a stuck tool loop ends cleanly with a Continue offer instead of burning the whole turn
Agent email sends route through the desktop's own channel plumbing behind a hard confirmation gate, so every outward send asks you first on every mode, and the IMAP email channel card shows its connected state properly
Projects: chats created in a project save to the project's shared workspace, a Project History tab shows the timeline, files support drag-and-drop moves and click preview, and chat rows show creation dates
Model picker: newest models first, embedding, image, and audio models filtered out of chat, and the ChatGPT subscription catalog pulls live from your connected account
Long sessions run leaner: smarter prompt caching and tool-catalog handling cut tool payloads by up to 86 percent
Smarter updates: differential downloads make updates smaller, auto-update waits while sessions, tasks, or teams are still working, and Windows explains exactly what it needs when an update requires elevation
Scroll up freely while the answer keeps streaming, enabled MCP connectors reach every agent backend, the Google Workspace connector got a major refresh with a clean browser consent hand-off, skill imports run a scan-and-verify security sweep with live progress, and the bundled engine updated to v0.12.25
v0.11.8June 30, 2026
New
Browser use, overhauled: Wayland bundles and auto-enables a built-in browser capability, so when a job touches the web it opens pages, clicks, fills forms, and pulls back exactly what you need, fast and clean
Computer use on the Mac, sharper and smoother: a one-pass permission onboarding flow gets screen recording and accessibility live in seconds, with deep-links and the usage strings macOS needs
Added
Response-length control: set the prompt response length to Auto or a fixed budget you dial in, and the engine sizes the right token budget per model automatically, so you get big complete answers when you want them and tight, lean ones when you do not
Persistent per-project workspace: projects keep a persistent workspace if you do not set a location, so everything you build stays exactly where you left it, session after session
Improved
Reliable channel delivery: named messaging channels, including an IMAP email channel, resolve and deliver correctly, so when you wire up a channel Wayland delivers every time
Massive projects run lean: large project context is capped so big projects stop ballooning token usage and cost, all the work at a sliver of the cost
Long sessions run clean across every provider: internal routing metadata is stripped from outbound tool calls, so long-context sessions stop getting rejected by strict providers
Concierge gained workspace and config-path diagnostics, and the bundled engine updated to v0.12.17
v0.11.7June 29, 2026
New
Concierge, a guide that lives inside Wayland: the first built-in, setup-aware helper that reads your own Wayland, the models you have connected, your skills, and your scheduled tasks, then answers anything you ask in short, plain English with real steps
Concierge sets things up for you: it proposes a change, you see a clean confirmation card and click Apply, and it connects an AI provider, sets your default model, adds a connected tool, or edits an assistant, with you in control the whole way. The first out-the-door v1, with big plans to expand it
Added
Suggested starting cards on every assistant, so you are never staring at a blank box, on a cleaner home screen
Draft recovery: your chat draft survives a close, reload, or restart and waits right where you left it
Improved
A steadier Flux Auto reasoning budget, so big, thinky replies stop getting cut off mid-thought, great for reasoning models
Smarter macOS auto-update detects and recovers if an update ever quietly gets stuck on an old version, and local MCP servers stay connected on macOS instead of dropping out
A connected Claude Code agent heals its own cache instead of refusing to start, and comes up cleanly
Telegram and Signal connectors optimized, OpenCode models show up where they should, the close icons over the Mac title bar are clickable, the chat menu stops clipping Delete, and the bundled engine updated to v0.12.16
v0.11.5June 28, 2026
Added
Keyless local models: local Ollama and llama.cpp endpoints need no provider key at all, so you point at a self-hosted endpoint and run fully private and free, with zero keys
Improved
A cleaner, tidier sidebar with a calmer, more focused layout
Bundled engine updated to v0.12.15, where keyless self-hosted endpoints land
v0.11.4June 28, 2026
Added
Crucible, a Mixture-of-Providers council in the bundled engine: it convenes multiple models from different providers on one task, lets each take a swing, and fuses the strongest answer
Flux Auto routes each request to the best model automatically, so you stop guessing which AI to use
Real tool-use on Windows: Wayland runs real shell commands and tools inside a locked-down AppContainer sandbox, on a clean non-developer machine, live-verified
Tools for local models: Ollama and llama.cpp run private and free and now handle tools, with up-front capability detection and a graceful no-tools fallback
One-click assistants with starter-prompt grids, multi-model teams, and scheduled tasks
Improved
Greatly improved chat streaming: an inline orbit indicator, a unified plain-English activity timeline across every backend, live reasoning summaries in the thinking block, smooth fast streaming, a per-message action toolbar the moment a turn completes, and an accurate turn timer even when you switch chats
Accurate live model catalogs for every connected provider, and assistants run on Wayland Core by default
Grok "Sign in with X" stays authenticated through a single race-free OAuth refresher
Cleaner billing-error classification, a hardened model picker that never strands you on an empty list, and the bundled engine updated to v0.12.14
v0.11.3June 24, 2026
Added
Sakana AI (Fugu) joins as a first-class provider, with its Fugu models and brand logo
Live model catalogs for ChatGPT subscription, MiniMax, and DashScope, streamed from each provider
Native ChatGPT subscription chats run through the engine's ChatGPT provider and are selectable directly
Local Ollama models appear in the picker and run keyless
Global Wayland Memory is wired into chat, so the assistant recalls you and your projects
Flux Voice gets its own dedicated configuration block
Improved
Every connected key routes straight to its own endpoint via the v2 registry bridge at spawn, across Sakana, OpenRouter, and the whole OpenAI-compatible family
Smarter model guidance: models without tool support point you to a tool-capable one, and Claude Code starts on Opus when no model is pinned
Cleaner approval prompts, contained file preview, and smoother back-after-error navigation
Sharper onboarding and Doctor, Windows npm and WSL CLI detection, MCP OAuth sign-in with timeout and cancel, and the bundled engine updated to v0.12.8
v0.11.2June 22, 2026
Added
Native routing for 104 providers, including xAI Grok, Perplexity, OpenRouter, Groq, DeepSeek, Cerebras, and NVIDIA
Flux Voice arrives as a speech-to-text provider so you can talk to Wayland
Pop-out windows: pop any main destination out of Mission Control into its own window
One-click migration brings provider keys and MCP servers over from Hermes and OpenClaw
GitHub Models tokens, OpenAI service-account and admin keys, and the new Google AI Studio key format are recognized
Improved
Redesigned provider picker leads with real brand logos and a bring-your-own-endpoint section up front
Friendly model names everywhere: the chat header and send box read Claude Opus 4.8 instead of a raw model id
Sharper Claude model switching: the picker holds your selection and switches the running model
Egress firewall is now a clear toggle in Security settings, plus smoother headless and server setup and automatic provider retries
v0.10.0June 16, 2026
New
MCP marketplace: browse and connect 107 curated integrations, or paste any MCP server URL
Image generation in chat with Flux, no separate key to set up
Runaway-loop circuit breaker: a budget gate and loop detector stop a session before it spins out
One-click Flux failover: reconnect and replay a broken chat without losing context
Custom MCP tools now flow into Claude Code and Codex sessions, not just built-in agents
Set up a provider key from your phone, plus 33 ready-made Operator Teams
Improved
Smarter skills discovery surfaces the right capability as you type
Security hardening: a multi-wave audit closed 80-plus findings
v0.9.7June 12, 2026
New
Sign in with X (Grok): one click signs you into your xAI account, no API key to paste
Browser-style tabs and pop-out windows for running conversations side by side
Update pill in the title bar, plus a real macOS app menu
Autopilot guardrail: a safety classifier catches catastrophic commands before they run
Author your own slash commands, plus a built-in Doctor health check
More provider connect flows: NVIDIA NIM, Ollama Cloud, Moonshot, with per-provider toggles
Improved
Full mobile and responsive pass: settings, composer, and every panel work on a phone
Cleaner providers screen with Flux Router featured up top
Fixed
Connection and routing bugs across Flux and the model picker, plus connect-panel overflow
v0.8Beta
New
80+ MCP integrations to connect your real stack
Bring-your-own-key model providers and on-the-fly model switching
Improved
Skills library expanded past 2,000 curated, audited skills
v0.7Beta
New
Scheduled tasks: set recurring work in natural language
Hands-free voice in and out
v0.12.25July 13, 2026Latest
Added
doc_extract, a new tool that reads office documents so agents can pull content straight out of Word, Excel, and PowerPoint files
Vision accepts local image files in vision_analyze, not just URLs, and images encode consistently across every provider
Per-assistant scoping for config MCP servers, so each assistant sees exactly the servers it should
Host-delegated email sends behind a hard per-send confirmation gate: every outward send asks first, on every mode, independently security-audited
A per-conversation autonomous-send rate cap, so two agents wired to the same channel can never loop forever
Isolated profiles serve over ACP, with a profile supervisor that runs one process per profile and a trusted persona roster
Improved
A major token-efficiency pass: smarter prompt caching, deterministic tool ordering, and a tool-catalog fold that cuts tool payloads by up to 86 percent
Per-model output sizing when no cap is set: known models get their real ceiling, so big answers stop truncating and small models stop erroring
Honest tool signals end to end: tool failures are visible to the agent instead of reading as success, a retry cap stops stuck tool loops cleanly, and hitting the per-turn cap offers Continue with fresh headroom
Consent hardening end to end: approvals carry a server-minted secret bound to the exact action, auto-approving modes are a launch-time opt-in, project config can only tighten posture, and untrusted project instructions are confined and neutralized
A stale-engine self-heal in the npx launcher that spots an old cached engine and prints the exact command to refresh it
Windows reliability: command execution unbroken across platforms, sandbox output pipes drain concurrently, and the AppContainer availability probe is wall-clock bounded
v0.12.17June 30, 2026
Added
Built-in browser capability bundled and auto-enabled, so web jobs open pages, click, and fill forms out of the box
Response-length control: a fixed budget you set or Auto, where the engine sizes the right token budget per model automatically
Persistent per-project workspace when no location is set, so project work stays put across sessions
Improved
Reliable named channel delivery, including an IMAP email channel that resolves and delivers correctly
Large project context is capped, so big projects stop ballooning token usage and cost
Internal routing metadata is stripped from outbound tool calls, so long-context sessions stop getting rejected by strict providers
v0.12.16June 29, 2026
Improved
The reasoning budget never starves the visible answer, so long, thinky replies finish their thought
Sandboxed bash falls back to cmd when PowerShell is selected under AppContainer, so shell tools stay reliable on Windows
Bounded WebFetch readability extraction with a per-turn breaker reset, for steadier web reads
Reasoning-content replay for strict reasoners routed via a router
v0.12.15June 28, 2026
Added
Keyless self-hosted endpoints: a local Ollama or llama.cpp endpoint runs with no provider key at all, so local models work fully private and free, with zero keys
Improved
A cleaner, tidier sidebar with a calmer, more focused layout
v0.12.14June 28, 2026
Added
Crucible (Mixture-of-Providers) council: N proposers each pinned to a different provider work in parallel, and a fenced read-only aggregator fuses one answer, with budget and daily-cap guards and 151 dedicated tests
Run it via wayland-core crucible "<task>" with --auto (a classifier convenes only when worth it), --advisor (inject the fused synthesis into the agent loop as private guidance), or --terminal
Capability-first tools gate: tool-incapable models degrade gracefully across local and cloud backends, and the choice is remembered per model
MCP smart tool-curation: a provider-aware cap keeps large MCP servers within the model's tool limit, with real server provenance and BM25 relevance curation
Provider-aware media tools: image generation and text-to-speech follow your active provider, with graceful fallbacks
Improved
Crucible reliability and cost accuracy: bring-your-own pricing catalogs load, Gemini pricing is accurate, and Opus support is broader in councils
Windows AppContainer shell tools work end-to-end on non-developer machines: real subprocesses run sandboxed, the filesystem allowlist skips absent dev caches, and the job tree is reaped so commands return promptly
v0.12.8June 24, 2026
Added
Sakana AI (Fugu) provider, OpenAI-compatible and ready to connect
Isolated profiles keep each CLI run cleanly separated
Active-window kernel sizes the context window accurately after a model swap
JSON-stream observability with active-window percentage, agent-run correlation, and structured traces
Improved
Smart auto-compaction at the active-window threshold, Flux-aware and with a clean memory handoff
A Flux context-routing contract plus tolerant handling of context overflow
OpenAI image generation defaults to gpt-image-1, and DeepSeek v4-flash brings 1M context
Windows reliability gains and a security dependency hardening batch
v0.12.6June 22, 2026
Added
Flux Router drives image generation, web fetch, and web search from the engine, with a new wcore image subcommand
MiniMax connects through an Anthropic-compatible endpoint, shows up in the model pickers, and lands in the cost catalog
One-command MCP Forge connect: a single /mcp connect closes the discover, authorize, and connect loop in the TUI
Improved
Leaner token spend with cheaper, accurate compaction, stable prompt cache prefixes, and a focused prompt for spawned subagents
Stronger workspace sandbox: a unified WorkspacePolicy roots the Bash sandbox at your workspace, with OS-level secret-read deny and Landlock on Linux
Sharper Windows support: a shell selector accepts pwsh.exe, powershell.exe, and absolute paths, with project context scoped to the project
More reliable provider auth: Grok OAuth token refresh, MiniMax region failover, and machine-readable auth codes
Smoother native Gemini chat that reaches its host under the default egress policy and surfaces a real message on every turn
Cleaner ChatGPT subscription model lists, plus wider egress allowlists for NVIDIA NIM, Cerebras, MiniMax, Qwen, and Flux Router out of the box
v0.12.0June 15, 2026
New
Two-way inbound channels: agents now receive messages and media across Telegram, Discord, Slack, WhatsApp, Matrix, Teams, and SMS
Native per-command Bash output compaction keeps long build logs out of the model's context
Plugin marketplace: natively install Claude Code marketplaces and plugins
Default runaway-loop circuit breaker with loop-convergence safeguards
Parallel free search by default, with a Firecrawl, Exa, and SearXNG fallback ladder
Improved
MCP boot resilience plus /doctor surfacing of connection issues
gpt-5 family routes to the OpenAI Responses API; international endpoints pinned for 8 native providers
v0.9Beta
New
ForgeFlows: declarative multi-agent workflows in a RON DSL
Cognitive memory with auto-indexing across sessions