Changelog
HexClaw Desktop engineering progress and release timeline. Continuous iterations for local-first autonomy.
v0.5.1#
Latest- ImprovedSolve and grade in regular chats: Send text or image questions to a regular assistant without creating a child profile first. When the question is legible but the student’s answer is unclear, it stays ungraded with a separate reference solution.
- ImprovedTeaching reports and PDFs: Solutions, grading results and guidance for parents come together in one report, with teaching diagrams where needed. Preview, download or print the matching PDF. Verified answers remain available if the explanation fails to generate.
- ImprovedWeekly and supplementary practice: Generate practice by section, add supplementary sheets and resume interrupted work. Retries keep the original question count; manual practice has its own count and does not use the automatic weekly allowance.
- ImprovedLearning profiles and progress: Save conflicts preserve your input; backup restoration keeps original images and verified question sources. Progress uses the actual submission time and subject, distinguishing independent work from answers completed with hints.
- ImprovedPreview and export chat results: Preview complete HTML results, inspect their source and download them. Markdown, JSON and PDF exports retain images uploaded in the conversation.
- ImprovedResume media tasks: Image, video and speech generation retain execution and delivery states for recovery after interruption. Historical media and local attachments load more reliably, with unknown outcomes shown in their actual state.
- ImprovedSkills and MCP: Installing, enabling, disabling or removing a Skill or MCP service updates the runtime. Skill bundles can include resources such as templates and scripts.
- ImprovedConnections and agents: Connections show usage and test status, retain agent bindings after renaming and preserve drafts when templates load. Process labels and Uyghur layouts are more consistent across the three languages.
- ImprovedLong-term memory: Relevant memories remain available with a Skill attached, scoped by agent, persona or business object. Response conventions are checked each turn. Background refresh keeps the list, search and filters in place.
- ImprovedLong conversations and sources: Conversation compaction preserves original messages and image references. Search retains source links and supports reading matched pages. Past answers keep the model source used for that turn.
v0.5.0
- ImprovedWorkflow canvas: Configure nodes, follow run progress and browse past runs. Import or export workflows, trigger them on a schedule or through a Webhook, and save results to a chosen local directory.
- ImprovedUnit study materials: Completed materials are saved to the library with a PDF matching the text. Reopen them to read the saved result, check textbook sources and see which teaching Skills were used.
- ImprovedMath checks and annotated homework: Programmatic checks supplement math solutions, with verified questions reused when source and curriculum conditions match. Correcting a source keeps the original question numbers and corresponding annotated images.
- ImprovedModels and native reasoning: Model settings distinguish native reasoning support from the thinking switch. Probe reasoning separately and set its effort level. Saving configuration preserves existing credentials and capability declarations.
- ImprovedConversations and attachments: Improve compatibility with historical attachments and the stability of startup and conversation switching. Tool activity is consolidated, and internal page markers are removed from retrieved knowledge text.
- ImprovedTeaching messages in DingTalk: Improve emphasis, links and summaries. Long teaching content supports a short summary with a complete PDF for reading on a phone.
- ImprovedMemory switches apply immediately: The automatic memory extraction and recall switch takes effect after saving, without a restart. Turning it off retains existing memories and manual management.
- NewHistorical log search: Search disk logs by time, level, source and keyword, with pagination and grouping. Records within the retention period remain available after a restart.
v0.5.0-beta.3
- NewK12 homework tutor: Added support for grades 1–6 in math, Chinese, English, science, information technology and art. Included separate assistants, grade and semester settings, math textbook editions and learning profiles for each child.
- NewTeaching guides: Added parent guides for all six subjects with full solutions, grade-appropriate methods, teaching steps, hints and questions to check understanding.
- NewPhoto grading: Added grading of completed homework with annotations on the original photo, plus support for full solutions and parent teaching guides for unanswered questions.
- NewWriting and art feedback: Added revision suggestions and art feedback tied to the original text or visible details, with reasons for each comment.
- NewLearning records: Added mistake records, learning notes, artwork and progress records. Supported reviews at 1 / 3 / 7 / 14 days, practice generation, printable worksheets and photo grading.
- NewExports and backups: Added PDF, Word and Markdown learning-profile exports with ZIP attachments, plus K12 backup and restore using
.hexbakfiles. - NewDingTalk tutoring: After binding a channel, send homework photos from a phone and receive explanations and annotations in one-to-one private chat.
- NewScenario packages: Separated scenario-specific views, data and APIs from general Agent capabilities, starting with the K12 package.
- ImprovedChat and integrations: Updated streaming chat, math rendering, knowledge-base pagination and source filters, MCP restarts and Ollama model management.
- ImprovedWindows document tools: Aligned Windows packaging with document-rendering dependencies.
v0.4.9
- ImprovedAgent editing: Updated the structured SOUL editor, Skill picker, advanced parameters and edit-lock states.
- ImprovedChat sessions: Added session branching, persisted thinking durations and a fallback for tool-only messages; revised delete and export actions.
- ImprovedAutomation and connections: Updated permission controls, schedule validation, MCP status refresh and per-instance channel binding.
- ImprovedEngine and startup: Added graceful engine shutdown and memory behavior settings, and embedded the splash logo.
v0.4.8
- ImprovedMessage order: Interleaved thinking, text and tool calls in causal order, with tool-call cards and a sub-agent panel.
- ImprovedMessage sending: Showed sent messages without waiting for the Auto-RAG knowledge lookup.
- ImprovedIM Agent binding: Bound each channel to an Agent, with model selection managed by that Agent.
- ImprovedKnowledge retrieval: Added visual RAG settings, date-range filters and pagination.
- NewContinuous tasks: Added long-running tasks that continue across execution cycles.
- ImprovedText-to-speech: Enabled desktop speech by default through free edge-tts, with interruption and retry controls.
- ImprovedWorkflow recovery: Added failure recovery that reuses completed nodes when resuming a workflow.
- ImprovedLong-term memory: Unified memory in FileMemory and moved it to the Long-term Memory page.
v0.4.7
- ImprovedReplies and feedback: Persisted error replies and like/dislike feedback across session switches and reloads, keeping feedback actionable.
- ImprovedSpeech feedback: Reported text-to-speech failures with a visible message.
- ImprovedOutput degeneration: Added detection and interruption of degenerate model output.
- ImprovedDocument previews: Added in-app download and preview, with a native save dialog and backend PDF parsing.
- ImprovedChat interactions: Adjusted scroll navigation in new sessions, prevented Enter from sending during IME composition and kept model selection within each session.
v0.4.6
- ImprovedMCP credentials: Encrypted stored MCP environment credentials with a master key and redacted them from API responses.
- ImprovedTool-call limits: Returned partial results with a continue hint at the turn limit. Raised the default to 25 turns, with a hard cap of 50.
- ImprovedNew memories: Loaded new memory entries without an engine restart.
- ImprovedNotification sources: Added cron and IM source tags for notification icons and deep links.
v0.4.5
- NewConnections hub: Added reusable IM and email connections with local encryption and domain-based SMTP/IMAP setup.
- NewData connectors: Added read-only GitHub and Notion access, connection checks and resource browsing, with server-side token encryption and redacted API responses.
- NewWorkflow editor: Added step creation, editing, reordering and removal, plus saving and dry runs with live node status.
- NewScheduled delivery: Added a picker for delivery targets from configured connections.
- NewAssistant persona: Added plain-text editing of
~/.hexclaw/SOUL.mdwith restore-to-default. - NewComposer shortcuts: Added the prompt library (
/,$ARGUMENTS), image drag/paste, the ⌘K command palette and an engine-disconnected banner. - ImprovedShared interface: Unified controls, styling and card hover effects, with Simplified Chinese, English and Uyghur (RTL) coverage.
- ImprovedWindows installation: Switched to NSIS (
.exe) installers and dropped.msi. Replaced bundled Ollama with a guided separate install. - ImprovedDisplay details: Updated the engine version display, removed overlapping trace labels in Logs and completed saving-state translations.
v0.4.2
- ImprovedNavigation: Gave Chat its own entry and grouped other pages under Build, Connect and System.
- ImprovedSkill details: Added a read-only
SKILL.mdview and standardized the Skills terminology. - ImprovedKnowledge documents: Added Markdown rendering, code wrapping, lighter code-block styling and hover-to-copy.
- ImprovedWebhooks and sessions: Updated disabled-webhook handling and navigation synchronization, with a consistent session write API contract.
- ImprovedDocument rendering: Moved pandoc and typst to Tauri sidecars with download integrity checks and made Windows packaging compatible with git-bash.
v0.4.1
- ImprovedCode rendering: Added code-view rendering and downloads with bundled pandoc and typst tools.
- ImprovedScheduling and streaming: Unified the cron scheduled-task architecture and included end-to-end SSE streaming compilation.
v0.4.0
- ImprovedUnified composer: Moved text, image and video requests into one ChatInput. Removed separate media composers and parameter panels; image and video generation used default parameters.
- ImprovedUyghur layouts: Added RTL support with translation keys mapped to Simplified Chinese. Kept mixed scripts correctly oriented and code blocks LTR.
- ImprovedInterface styling: Adjusted borders, shadows, transitions and accent glow, and adopted the Apple system font stack.
- ImprovedMedia preview: Added a persistent image/video download button, RTL-aware positioning and full-screen previews.
- ImprovedLocal media: Saved generated outputs to
generated/and preferred non-expiring local URLs to reduce base64 data in the database. - NewInteractive messages: Added Buttons, Select, Approval, Card and Block components with the
message.interactiveprotocol. - ImprovedRuntime capabilities: Added feature flags, Skill Pipeline, model-capability probing, event transport v1, model gateway v1 and tool lifecycle v2.
v0.3.8
- NewImage generation: Added ZhipuAI CogView-4 with embedded data URI images and an ai-core ImageProvider interface.
- NewVideo generation: Added ZhipuAI CogVideoX-2 with async polling, embedded covers and inline playback, plus an ai-core VideoProvider interface.
- ImprovedTool names: Localized 15 built-in tool names while keeping raw names for unknown MCP and custom tools.
- ImprovedModel selection: Kept normal chat model selection separate from default Agent preferences and showed Agent attribution in Agent mode.
- ImprovedMultiline memories: Kept multiline memories as single entries and supported updating, deleting and archiving them.
- ImprovedOllama downloads: Adjusted timeout handling for large-model downloads and localized download status.
- ImprovedRouting controls: Hid the strategy selector when smart routing is disabled.
- ImprovedFeishu reactions: Replaced social emoji reactions with Typing and THINKING indicators.
v0.3.7
- ImprovedThinking display: Persisted thinking durations and added a timed, expandable reasoning view with a left border.
- NewSession titles: Generated titles from the first user message without a model call.
- ImprovedMessage details: Persisted provider, model and Agent details across restarts, with time, model and Agent in one footer row. Removed assistant bubbles and made message width adaptive.
- ImprovedSandbox network: Applied changes to sandbox network settings at runtime.
- ImprovedFile tools: Restricted
file_edit,grepandglobto the workspace, including symlink escape checks. - ImprovedTool errors: Returned full error reasons so the model could correct its next attempt.
- ImprovedContext compression: Added model-based and heuristic summaries that retain user constraints.
- ImprovedGemma 4 thinking: Added thinking timeout protection without injecting
/no_think.
v0.3.6
- ImprovedStreaming scroll: Allowed scrolling upward during a response without auto-scroll pulling the view back down.
- ImprovedAgent and IM controls: Updated the Agent selector and IM channel configuration interface.
- FixedNew sessions: Fixed the first message failing to send in a new session.
v0.3.5
- ImprovedOllama warmup: Loaded the first downloaded model at startup to reduce the wait for the first chat.
- ImprovedHistorical reasoning: Handled
<think>tags when loading historical messages to keep reasoning separate from answer text. - ImprovedMCP home directory: Added the user's home directory when installing the filesystem MCP server on macOS, Windows and Linux.
v0.3.4
- ImprovedVersion display: Waited for the engine to start before fetching its version and synchronized app version values across the Sidebar, About and package files.
v0.3.3
- ImprovedVersion display: Used the Tauri app version in Sidebar, About and System Status, and added
engine_versionto the version API type. - ImprovedMCP home directory: Added cross-platform home-directory detection for the filesystem MCP server.
v0.3.2
- ImprovedReasoning extraction: Separated reasoning embedded in
<think>tags from message text for models such as glm-z1-airx. - ImprovedSettings dialogs: Adjusted dialog width to fit the content.
- ImprovedEngine restart: Added automatic retries when restarting the sidecar failed.
v0.3.1
- ImprovedModel selector: Replaced the dropdown with model chips for viewing and switching configured models.
- NewModel discovery: Fetched available model lists from cloud providers.
v0.3.0
- ImprovedMCP tools: Completed the marketplace tool-call flow and returned tool error messages.
- ImprovedSkill triggers: Refined keyword matching for Skill triggers.
- ImprovedOllama models: Loaded models into memory after startup and added download progress and running-state indicators.
v0.2.6
- NewRequest tracing: Added trace IDs across WebSocket, gateway, engine and LLM calls, with IDs and structured fields visible in Logs.
- FixedSession messages: Fixed a deadlock that blocked consecutive messages in the same session.
- ImprovedCloud MCP calls: Preserved MCP tool calls in cloud model streaming responses.
- NewThinking toggle: Added a thinking control in the composer, with
/no_thinkinjection for local thinking models. - ImprovedStream cancellation: Made stopping a response cancel its backend request.
v0.2.5
- NewReasoning display: Added streaming reasoning for Qwen3 and DeepSeek-R1, with automatic expansion and collapse, a timer and persistence.
- ImprovedOllama chat: Updated local-model listing, WebSocket authentication and streaming timeout handling.
- ImprovedStartup screen: Matched the system's light or dark appearance and waited for the engine to be ready.
- ImprovedKnowledge uploads: Closed the upload dialog after file selection.
- ImprovedMessage editing: Updated edit-button translations and sizing.
- ImprovedInterface controls: Removed the unavailable System Status tab and microphone button; the Tauri WebView lacked the Web Speech API.
- FixedHomebrew startup: Fixed first-chat errors caused by a sidecar version mismatch and a nil embedder.
v0.2.4
- NewMCP marketplace: Added Skill/MCP filters and one-click MCP server installation from
mcp-registry.json. - ImprovedMCP connections: Updated connected-state display and enabled MCP without a preconfigured server list, allowing servers to be added later.
- ImprovedmacOS tool paths: Added Homebrew, nvm and Cargo paths to the sidecar environment when launched from the macOS GUI.
- ImprovedSecurity controls: Tightened CSP, added an MCP command allowlist and knowledge-upload zip-bomb checks, and made configuration writes atomic.
v0.2.2
- NewModel providers: Added provider logos and support for Zhipu AI, Kimi (Moonshot), Wenxin Yiyan (Baidu), Tencent Hunyuan, iFlytek Spark and MiniMax. Added keyboard navigation, ARIA support and viewport-aware menu positioning.
- NewOllama installation: Added detection, installation, status polling and model association in one flow.
- ImprovedMCP tabs: Kept MCP tab states mutually exclusive and completed translations.
- NewmacOS installation: Added a one-line installer and Homebrew Cask support.
v0.1.0-beta
- NewDesktop beta: Released beta installers with first-run onboarding and default role setup.
- ImprovedModel providers: Expanded support for OpenAI, DeepSeek, Anthropic, Gemini, Qwen, Doubao, Ollama and custom OpenAI-compatible endpoints.
- ImprovedChat workspace: Expanded multi-model switching, artifact previews, document parsing, export and deep research mode.
- ImprovedTools and automation: Expanded MCP tools, RAG knowledge bases, semantic memory, scheduled jobs, the ClawHub Skill marketplace, IM channels and security controls.
v0.0.1
- NewMulti-model chat: Introduced chat with OpenAI, DeepSeek, Anthropic, Gemini, Qwen and Ollama.
- NewAgent workflows: Added Agent orchestration, multi-Agent meetings and a visual workflow canvas.
- NewMCP tools: Added MCP protocol integration.
- NewKnowledge and memory: Added RAG knowledge bases and semantic memory.
- NewScheduled tasks: Added scheduled task dispatch.
- NewSecurity gateway: Added the security gateway.
- NewIM channels: Added Feishu, DingTalk, WeCom, WeChat, Slack, Discord and Telegram connections.
- NewDeep research: Added deep research mode.
- NewDesktop platforms: Released native support for macOS, Windows and Linux.
Ready to experience HexClaw Desktop?
Local-first, private and secure, multi-model orchestration. Get started on macOS, Windows, or Linux.