Repository navigation
Releases
v2.10.0
RemiAI v2.10.0 Release Changelog
In a real RemiAI deep-research follow-up using the same model, input usage dropped 93%, from 524,663 to 36,480 tokens. RemiAI sent 80% fewer tool schemas per call and used 53% fewer input tokens on the first call, before any new tool results came back.
Added
- Adjust chat text size from 85% to 150% in Appearance settings. Changes preview immediately, apply to messages and the composer, and persist across sessions.
- Create interactive calculators, dashboards, charts, maps, tables, and media widgets directly in chat. Bar charts support validated data and responsive rendering.
- See input and output token counts beside the response completion time. Hover for exact counts.
Improved
- Follow progress in the Work activity timeline. Expand individual tool actions to see details with sensitive values redacted. Consecutive successful web fetches appear in one collapsible entry, and interrupted work is marked after a disconnect.
- Open sources and file changes together in the expandable response details. Goal progress and Build history use a shared responsive side panel, with improved focus handling and session-file links.
- Reduce model input during tool-heavy conversations. RemiAI compacts older tool results, retrieves archived details on demand, and extracts readable content from HTML pages. In one same-model deep-research follow-up reported in the commit, input usage fell 93%, from 524,663 to 36,480 tokens across the response. That comparison also recorded 80% fewer tool schemas per call and 53% fewer input tokens on the first call.
Fixed
- Keep Plan mode active until explicit approval. Reconnect to the correct stream and recover automatic replies more reliably after a disconnect.
- Keep session-file and canvas views synchronized after manual changes, and preserve session-file context when opening panels.
- Validate SVGs before rendering, preserve complete SVG output, and repair saved bars that were drawn with negative rectangle dimensions.
Maintenance
- Bumped the application and lockfile to 2.10.0 and added migration
0056_chat_text_size.sqlto store the new preference. - Expanded automated coverage for stream recovery, activity persistence, visual rendering, and live tool context.
- Removed the macOS installation notice from the installer workflow. The project TODO still tracks macOS notarization and now includes connecting the desktop app to a remote deployment.
v2.9.0
Added
- Added project workspaces: create and organize projects, link chats, attach project files, pin and reorder projects, and start chats with project-aware context.
- Added the cross-chat Library, making generated and attached files easier to browse from a dedicated sidebar destination.
- Added Guided Work mode with durable work runs, an explicit planning and approval lifecycle, target selection for directories or canvases, progress events, changed-file evidence, and verification results.
- Added durable chat-generation runs, replayable streams, server-sent lifecycle events, and generation-presence indicators so active work remains visible and recoverable across clients.
- Added a License settings page, application metadata, request-duration tracking, and refined token-usage visualizations.
Fixed
- Fixed stale generation-presence state and improved background-generation status and notification behavior.
- Improved chat-history search-context compaction to keep relevant context within limits.
- Improved chat scrolling, sidebar action visibility, mobile layouts, file picking, and settings usability.
- Hardened macOS installer preparation and Docker-release registry caching.
Changed
- Updated the application and lockfile version to 2.9.0.
- Added migrations for chat generation runs, projects, and Guided Work runs.
- Expanded automated coverage for projects, generation lifecycle and presence, stream recovery, search-context compaction, persistence intervals, and question delivery.
macOS installation notice
These macOS installers use an ad-hoc signature and are not notarized by Apple. macOS will require app-specific approval before opening RemiAI.
- Download the DMG for your Mac:
arm64for Apple Silicon orx64for Intel. Download its matching.dmg.sha256file too. - In Terminal, go to the folder containing both files and run
shasum -a 256 -c RemiAI-<version>-mac-<arch>.dmg.sha256, substituting the release version and your architecture. It should reportOK. If the checksum differs, delete the download and get a fresh copy from this release; do not open the mismatched file. ZIP downloads have their own.zip.sha256files. - Open the DMG and drag RemiAI to Applications. Try opening RemiAI once. If macOS blocks it, open System Settings → Privacy & Security, scroll to Security, and click Open Anyway for RemiAI, then confirm Open. Apple shows this option for a limited time after an opening attempt.
If macOS still says the app is damaged after checksum verification and app-specific approval, report the macOS version, selected architecture, and exact alert in an issue. Do not disable Gatekeeper system-wide.
v2.8.1
macOS installation notice
The macOS installers in this release are currently unsigned and not notarized by Apple. macOS Gatekeeper may show a warning or prevent the app from opening normally. Only bypass that warning if you downloaded the installer directly from this official release and trust it.
v2.8.0
macOS installation notice
The macOS installers in this release are currently unsigned and not notarized by Apple. macOS Gatekeeper may show a warning or prevent the app from opening normally. Only bypass that warning if you downloaded the installer directly from this official release and trust it.
v2.7.0
RemiAI v2.7.0 Changelog
Release Notes
Added
- Added the autonomous Heartbeat system with scheduled runs, configurable providers/models, tool-group and tool-name allowlists/denylists, retry limits, retention, and completion notifications.
- Added durable automation-run tracking, run controls, heartbeat tool-call history, and related API/settings surfaces.
- Added categorized memories with event dates, editing, import support, and memory-management UI.
- Added push-notification registration, streaming, and test support.
- Added RemiAPI visual-card support and configurable card display modes.
- Added HTTP request tooling and unified request-access controls.
- Added talk-mode voice pipeline improvements, microphone access handling, TTS, and transcript UI.
- Added model-preference controls for newly discovered provider models.
Improved
- Improved chat retry/recovery behavior for unanswered turns.
- Improved provider/model selection, backup/export/import flows, notifications, profile settings, and general UI presentation.
- Improved tool routing, web-search fallbacks, canvas handling, skills, session files, and automation quality policies.
- Updated the Next.js/Turbopack configuration and runtime path handling.
- Updated Docker/demo and RemiAPI worker documentation and configuration.
Fixed
- Fixed heartbeat provider fallback behavior and restricted heartbeat tool access.
- Fixed backup, memory, conversation, notification, media, and talk-mode issues.
- Fixed assorted Windows, packaging, rendering, and reload-loop issues.
v2.6.1
Merge branch 'main' of https://github.com/Houloude9IOfficial/RemiAI
v2.6.0
Added
- Added a public Demo button to the marketing website.
- Added a safe Docker public-demo mode.
- Added slash-command tool routing.
- Added configurable reasoning-effort controls across supported providers.
- Added command support and a connected chain-of-thought UI rail.
Improved
- Improved chat input and follow-up suggestion interfaces.
- Improved general application UI and the 404 page.
- Improved Windows compatibility.
- Improved demo imagery and marketing-site presentation.
- Improved lint configuration by excluding generated and user-created content from linting.
- Downgraded explicit
anyusage to warnings while retaining visibility.
Fixed
- Fixed middleware behavior.
- Fixed multiple UI-side issues.
- Fixed low-risk lint issues, including unnecessary
letdeclarations, calendar children handling, and intentional CommonJS imports in the Electron packaging hook.
v2.5.0
Overview
RemiAI v2.5.0 introduces the Canvas feature — an interactive, multi-file web project panel with live preview and code editor — alongside Docker multi-architecture builds, backup export robustness, and Anthropic provider discovery fixes.
This release contains 4 commits since the v2.4.0 release branch. No database migrations are required.
Highlights
Canvas — Interactive Web Projects
- New Canvas system lets the AI create, manage, and present self-contained multi-file web projects (HTML/CSS/JS) inside an interactive panel with live preview and a built-in CodeMirror code editor.
- Four new tools:
canvas_create,canvas_add_file,canvas_list, andcanvas_open. - Canvases live in the conversation's session sandbox under
canvas/{slug}/, inheriting all existing session-file infrastructure (MIME serving, real-time SSE change stream, backup/export). - A
canvas.jsonmanifest records name, description, and entry file — no new DB table needed. - Canvas tools are loaded on demand via the intent classifier (keywords: "interactive", "website", "game", "calculator", "prototype", etc.) and automatically activate
session_fileswhen triggered. - New UI components:
CanvasPanel.tsx(full editor + preview),CanvasPresentCard.tsx(inline result card with auto-open), and canvas-specific CSP/X-Frame-Options headers for safe iframe preview. - Canvas download route for exporting canvas files.
- Canvas test script (
scripts/test-canvas.ts).
Docker Multi-Architecture Builds
- The Docker release workflow (
.github/workflows/docker-release.yml) now builds for both AMD64 and ARM64 via QEMU + Buildx. - Added
docker/setup-qemu-action@v3andplatforms: linux/amd64,linux/arm64to the build step.
Backup Export Robustness
- Replaced the fragile
unwrap()helper inlib/api/backup.tswith a defensiveparseJson()that reads the response body as text first, attempts JSON parsing, and produces clear, actionable error messages when the server returns non-JSON (e.g. HTML error pages from proxies or load balancers). - Fixes the cryptic
JSON.parse: unexpected character at line 1 column 1error some deployments encountered during backup export/import.
Anthropic Provider Discovery Fix
- Fixed Anthropic model discovery to use
max_input_tokens(Anthropic's actual API field) as the primary source for context-window size, falling back tocontext_windowfor proxy re-maps. - Added explanatory help text in the Anthropic provider form clarifying that identity-linked (workspace-scoped) API keys require an
anthropic-workspace-idheader and won't return models — directing users to create a fresh "unlinked" key.
Changelog
Added
- Canvas feature: interactive multi-file web projects with live preview + code editor panel.
- Canvas tools:
canvas_create,canvas_add_file,canvas_list,canvas_open. - Canvas storage layer built on session-file sandbox (
canvas/{slug}/). - Canvas present card with auto-open behavior.
- Canvas download route.
- Canvas API endpoint (
GET /api/canvases). - Canvas-specific CSP and X-Frame-Options headers for safe iframe preview.
- Canvas intent classifier group with keyword matching.
- Canvas test script (
scripts/test-canvas.ts). - Docker multi-architecture builds (AMD64 + ARM64).
Changed
- Backup API client uses defensive JSON parsing with clear error messages for non-JSON responses.
- Anthropic model discovery reads
max_input_tokensas primary context-window field. - Refresh error message when auto-discovery is unavailable now says "models unchanged" instead of "no default models exist".
Fixed
- Anthropic models not showing up due to incorrect context-window field mapping.
- Backup export/import crashing with cryptic JSON parse errors on deployments with proxies or non-JSON error responses.
- Anthropic identity-linked (workspace-scoped) API keys silently returning no models — now documented in the provider form.
v2.4.0
Overview
RemiAI v2.4.0 is a focused release that adds ChatGPT-style temporary chats and per-chat memory toggling, expands native provider support to Google (Gemini), Mistral, OpenRouter, and Groq, adds automatic 5-minute model refresh and broader model discovery, improves citation and source-evidence presentation, and fixes desktop/app launch and tooling issues. It includes a complete redesign of the marketing website with new visual assets.
This release contains 10 commits since the previous release branch, including database migrations 0039–0040 and updates across the chat runtime, provider system, research presentation, desktop packaging, and the website.
Highlights
Temporary chats and per-chat memory toggling
- Added a ChatGPT-style personalization menu on each conversation with two independent switches.
- Temporary chat marks a conversation as "temporary": it gets a dashed visual treatment and is automatically deleted after 3 days of inactivity.
- Memory toggle per chat: turning memory off means the AI sees no saved memories, profile, preferences, or file access for that conversation and cannot save anything new.
- Temporary chats remain temporary when duplicated, and the memory toggle carries over too.
- Added server-side cleanup of expired temporary chats (with their uploads and session files) that runs once at server boot and is idempotent.
- New migration
0040addsis_temporaryandmemory_enabledcolumns to conversations.
Native provider support expansion
- Added first-class native support for Google (Gemini), Mistral, OpenRouter, and Groq via their official SDKs.
- New dependencies:
@ai-sdk/google,@ai-sdk/groq,@ai-sdk/mistral, and@openrouter/ai-sdk-provider. - Model discovery additions for the new providers, including Gemini context-window (input token limit) reporting.
- New migration
0039addsprovider_models.context_windowso the real context-window reported by a provider's models API is captured (falling back to hardcoded heuristics when unavailable).
Automatic model refresh
- Added a 5-minute auto-refresh of provider models so newly released models appear without a manual refresh.
- Refreshing is now respectful of deleted models (deleted models are not resurrected on every refresh), and the refresh endpoint/manual flow was consolidated.
- Improved the provider model list experience with clearer capability and compatibility handling.
Citations and source evidence
- Reworked the source evidence card and markdown citation rendering for a richer citation interaction, including a modal on hover to inspect citation/source details.
- Improved inline citation pills, evidence grouping, and the aggregated source evidence card.
- Tightened citation-related streaming and cleanup plumbing.
Stability, fixes, and reliability
- Fixed the Electron app not launching (desktop startup/auto-updater path).
- Fixed tool calling and streaming issues, including tool-group loading labels and message-stream cleanup.
- Fixed buggy ChatInput animations.
- Fixed code-exec tooling and tool-groups edge cases.
- Refactored the system-prompt builder with a verification script (
scripts/verify-prompt-identity.mjs) that guarantees byte-identical output with the previous implementation.
Website redesign
- Full redesign of the marketing site (
website/): reworked homepage, layout, header, dashboard shot, copy button, FAQ, and OG/social cards. - New brand assets
RemiAI-Dark-2-4-0.pngandRemiAI-Light-2-4-0.png.
Database migrations
The release includes migrations for all schema additions introduced since the previous release:
0039_provider_models_context_window.sql—provider_models.context_window0040_temporary_chats.sql—conversations.is_temporaryandconversations.memory_enabled
Multi-statement SQLite migrations use Drizzle's --> statement-breakpoint separator.
Changelog
Added
- Temporary chats with 3-day auto-delete and per-chat memory toggling via a personalization menu.
- Native provider support for Google (Gemini), Mistral, OpenRouter, and Groq.
- 5-minute automatic model refresh with model discovery and deleted-model preservation.
- Provider context-window capture and provider model-list capability improvements.
- Richer citation/source presentation with a modal-on-hover evidence interaction.
- System-prompt refactor verification script.
- New database migrations
0039–0040.
Changed
- Reworked the source evidence card and markdown citation rendering.
- Consolidated and improved the model refresh flow and provider model list.
- Redesigned the marketing website and updated brand assets to 2-4-0.
- Refactored the system-prompt builder (byte-identical output, verified).
Fixed
- Electron app not launching.
- Tool calling and streaming issues, including tool-group loading labels.
- Buggy ChatInput animations.
- Code-exec tooling and tool-groups edge cases.
- Citation cleanup and source-evidence styling.
- Lint errors: removed the
require()in the Electron auto-updater (converted to a staticimport) and rewrote thetoggleSelectAllmemoization inConversationList.tsxso React Compiler can optimize it.
v2.3.0
Overview
RemiAI v2.3.0 is a major capability release focused on reliable autonomous work, traceable research, durable automation, and a cleaner chat experience. It adds build-mode workflows with verification and repair, durable run tracking, artifact and source provenance, improved model routing and reasoning support, richer provider/MCP integrations, and a broad set of chat and settings UX improvements.
This release contains 29 commits since the previous release branch, including database migrations 0030–0038 and updates across the chat runtime, tool system, automation engine, research pipeline, provider settings, and desktop packaging support.
Highlights
Verified Build Mode
- Added a dedicated Build conversation mode for software tasks.
- Build runs track the task, definition of done, changed files, verification checks, checkpoints, result artifacts, summaries, and failures.
- Added build verification, quality policies, deterministic model routing, repair guidance, and bounded repair loops.
- Tool output now presents grouped file changes and verification results instead of flooding the conversation with individual tool cards.
- Added build-run history and APIs so completed and failed work remains inspectable after the response ends.
Durable automation and run control
- Added durable automation runs for routines, scheduled tasks, webhooks, and spawned agents.
- Runs track lifecycle state, attempts, checkpoints, retries, controls, events, errors, and completion times.
- Added stop, retry, and steer controls through run-control APIs.
- Added in-app notifications and a notification stream for background work.
- Existing routine, scheduled-task, webhook, and agent execution paths now share durable run tracking and retry behavior.
Artifacts, research, and provenance
- Added chat-scoped artifacts with metadata, versions, status, session-file paths, and result storage.
- Added durable research sources with URL, title, publisher, retrieval time, content hash, extraction status, freshness, and quality metadata.
- Added source claims with support states such as supported, partial, unsupported, disputed, and inference.
- Added citation extraction and inline citation pills, plus an aggregated source evidence card.
- Added research report, source, and artifact APIs so evidence can be reviewed independently of the chat transcript.
- Web research and fetch results now retain sanitized provenance without persisting full page bodies.
Model routing and reasoning
- Added quality policies: Fast, Balanced, Quality, and Selected.
- Added task-complexity estimation and deterministic routing across enabled providers and models.
- Added provider/model metadata to run traces and quality decisions.
- Improved reasoning support for Ollama and compatible providers, including conversion of reasoning deltas into
<think>streams. - Reasoning blocks are merged across multi-step responses and collapse when the final answer begins streaming.
Tool loading and agent workflows
- Continued the dynamic tool-loading system so only relevant tool groups are sent to the model by default.
- Tool groups can be activated by intent, recent usage, stored state, or the model during a response.
- Quiet bookkeeping tools are hidden when they run alone, while meaningful tool sequences remain grouped and visible.
- Improved sub-agent chaining, background execution, run tracking, and result retrieval.
- Added notifications and stronger continuation/repair behavior for long-running work.
Chat and attachment UX
- Redesigned the model selector with provider-aware model browsing and capability information.
- Redesigned the chat composer, model/capability controls, questions UI, attachment previews, and tool-call cards.
- Added improved image attachments, inline thumbnails, expandable previews, image-search galleries, and clearer file cards.
- Added citation pills and a collapsed-by-default source evidence card.
- Improved Markdown rendering, syntax-highlighted code blocks, copy actions, safe raw-HTML handling, and streaming animations.
- Combined reasoning sections and tool runs to reduce visual noise during multi-step responses.
- Added clearer model-selection and chat-start behavior, updated disclaimers, and improved mobile/sidebar layouts.
Providers, integrations, and settings
- Improved provider and provider-model management, including model discovery and compatibility handling.
- Added better reasoning capability detection across supported providers.
- Added JSON import for MCP server configurations.
- Added Brave image search with inline thumbnails and a gallery view.
- Added an update checker backed by GitHub releases and a sidebar update indicator.
- Added customizable background color support alongside existing appearance preferences.
- Improved profile, provider, MCP, and settings pages with clearer forms and more useful empty/loading states.
Database migrations
The release includes migrations for all schema additions introduced since the previous release:
0030_artifacts.sql— chat-scoped artifacts0031_sources.sql— durable research sources0032_research_claims.sql— source claims and source quality metadata0033_build_runs.sql— build-run persistence0034_quality_policy.sql— per-conversation quality policy0035_build_checkpoints.sql— build checkpoints0036_build_result_artifacts.sql— build result artifact references0037_durable_automation.sql— durable automation runs, events, and run links0038_background_color.sql— profile background color
Multi-statement SQLite migrations use Drizzle's --> statement-breakpoint separator.