Omnigent changelog: what's new each month
Every stable Omnigent release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 6 months covered; the current month updates daily.
November 2026 · so far
2 releases: 0.21.0 → 0.22.0
November brought major improvements to Omnigent's interface and automation capabilities, with better PR management, clearer agent actions, and expanded deployment options. The month also introduced nested agents, new automation triggers, and enhanced desktop features.
- omnigent opencode no longer supports OpenCode 1.x (opencode-ai); upgrade to OpenCode 2.0.18 or later.
- Desktop app requires macOS 13 Ventura or later; macOS 12 Monterey is no longer supported.
Highlights
- Pull requests now show in a compact list with status checks and merge actions, staying visible when the host goes offline.
- Agent actions display as readable plain-language summaries with inline results instead of raw JSON.
- Run sessions on any machine over SSH directly from the desktop app, with localhost apps opening in the browser pane.
- Automations can now be triggered by Schedule, Jira, Slack, GitHub, Linear, or PagerDuty from a single form.
- Nested agents organize spawned agents in the left sidebar for better clarity.
New
- Side chats have independent model and reasoning-effort controls for different follow-up configurations.
- Native desktop app menu and full keyboard shortcuts for terminals, tabs, pins, archive, and workspace navigation.
- Tag sessions, sub-agents, and files with @ in the composer.
- Pull request Changes tab can filter files, search diffs with Cmd/Ctrl+F, mark files as viewed, and browse individual commit diffs.
- AWS Lambda MicroVMs sandbox provider for running sessions.
- Sandbox traffic can route through upstream HTTP proxies.
- Edit sent messages back into the composer for correction.
- Agent replies can display screenshots saved in the workspace.
Improved
- Deployments can choose which built-in agents appear in New Chat and set a default starting agent.
- Hide individual harnesses and agents in Settings › Appearance.
- Model lists include search for easier navigation.
- Messages to Claude Code sessions are more reliable with improved handling of input box state.
- Chat automatically recovers when its live connection stops sending updates.
- Sessions wake automatically when background tasks or sub-agents finish.
- Sandbox and managed sessions can fork to any host or sandbox provider.
- omnigent opencode runs on OpenCode 2.0 with live text streaming, question forms, and native /compact support.
Fixed
- Messages no longer fail or stall when another program writes to the input box.
- Pull request linked sessions now stay visible while their host is offline.
- Chat recovery improved for silent connection drops.
October 2026
5 releases: 0.17.0 → 0.20.0
October brought major improvements to session reliability, pull request management, and agent customization. The month focused on surviving infrastructure disruptions, expanding version control integrations, and making the interface more responsive.
- POST /v1/sessions/{id}/switch-agent now returns 410 Gone; clients must fork the session into another agent instead of switching in place.
Highlights
- Sessions now survive server restarts, runner drops, and handoffs without showing stale status messages.
- Review and manage GitLab merge requests and Azure DevOps pull requests directly in the session panel, alongside GitHub.
- Install and manage reusable custom agents with new omnigent agent commands.
- Handle pull requests without leaving the Pull Requests tab: merge, edit titles and descriptions, manage labels and reviewers.
- Databricks Sandboxes can now be managed hosts, created and run as the user through their Databricks account connection.
New
- OIDC sign-in through your system browser in the desktop app with silent renewal and public PKCE clients.
- Inbox with All/Unread/Awaiting response tabs to catch up on agent replies.
- MCP registry and gateway support: connect managed MCP accounts and pick their tools for local hosts or sandboxes.
- Set reasoning effort for Claude SDK and Codex SDK agents from the web composer.
- Sandboxed agents can use approved Git repositories over SSH without receiving the SSH key.
- Type /mcp in the composer to see a session's MCP servers and connection status.
- Composer shortcuts: /clear for web chat in Claude Code and Codex sessions, Shift+Tab to cycle permission mode, ?prompt= URL parameter.
- Messages from the web now reach Claude Code sessions reliably with full support for long or multi-line prompts.
Improved
- Keyboard shortcuts for workspace tabs, common views, and recent sessions; Ctrl+Shift+L to focus chat input.
- Sub-agent status reports from confirmed outcomes prevent premature cancellations and false completions.
- Large PR diffs, sidebar query-cache scans, and source files beyond 2,000 lines no longer freeze the app.
- Pull Requests panel now shows pending reviewers, review states, and separates comments into their own tab with counts.
- Shell calls display as terminal output ($ command, output, exit status) and file links support line ranges like file.ts:10-20.
- Turns on offline laptops wait up to 24 hours for reconnection instead of failing immediately.
- Runner tunnels recover silently: stalled tunnels reconnect and requests keep processing during recovery.
- Claude Code auto mode no longer gets stuck on the environment-setup prompt.
Fixed
- Reconnected hosts no longer leave sessions marked disconnected after server restarts.
- Late events from a previous runner no longer cause session failures.
- Plugin slash commands and /btw questions sent from the web no longer report false delivery failures.
- Codex sessions recover after app-server stops, in-place Omnigent upgrades, or oversized tool calls in saved history.
- Claude Code sessions keep their compacted context after an Omnigent restart without repeating completion notifications.
- Native Codex --add-dir launches now work on Codex 0.156 and later.
September 2026
5 releases: 0.12.0 → 0.16.0
September brought major improvements to multi-repository workflows, native agent integrations, and session resilience. The month focused on making Omnigent faster, more reliable across network disruptions, and better at managing complex development projects.
- Bare omni command now starts the background host instead of launching the polly agent; use omni run for default agent.
- Web terminals moved to single control-mode attach path; legacy PTY transport configuration removed.
- Managed terminals now require tmux 3.3 or newer; warns during setup and fails early on older versions.
- Standalone omnigent-canvas extension removed; enable built-in Canvas with OMNIGENTFEATURES=canvas.
Highlights
- Devin now runs as a native harness, mirroring its TUI directly into the web UI like Claude Code and Codex.
- GitHub panel in the workspace shows PR summaries, comments, CI checks, and diffs; work across multiple repositories in one session.
- Canvas is now built-in with draggable session cards, live status, and branch tracking per project.
- Sessions recover automatically from server restarts and brief network outages without losing state.
- New managed-sandbox providers including libkrun microvms with sub-second boots and Kubernetes support.
New
- Import existing local CLI sessions (Claude Code, Codex, others) from the web UI with original titles preserved.
- Projects now drive new sessions, inheriting agent, workspace, and git defaults.
- Message deeplinks let you copy links to any message and jump to cited file lines.
- Side chats within sessions share context without polluting the main conversation.
- Curated model catalogs for Pi, custom ACP agents, and managed sandbox sessions with per-harness gateways.
- Run Omnigent host as a per-user macOS or Linux service with omni host enable/disable.
- Managed terminals using copyonwrite: true for disposable edits in sandbox writepaths.
- Admin controls for default session visibility, subdomain trust, and Slack setup pre-selection.
Improved
- Mobile and wide-screen UI redesigned with ChatGPT-style drawer, full-screen search, and better ultra-wide support.
- Chat composer redesigned to configure workspace, agent, model, effort, and permissions more smoothly.
- Sessions open immediately while server catches up; long transcripts virtualize for better performance.
- Claude context-window usage now visible; /compact command supports native compaction.
- Codex sessions resume more reliably, preserving conversation history and permission settings.
- Chat more accurately shows agent working state and keeps completed answers visible during background tasks.
- Unified workspace browser with compact file panel showing added, modified, and deleted badges.
- Brief runner drops no longer fail turns; runners reconnect promptly and transparently.
Fixed
- Pending approvals and user input requests survive server or runner restarts.
- Codex terminals recover without backend restart and survive launcher sign-in prompts.
- Model rate-limit errors can be retried directly from chat.
- Host launches no longer block on capability checks due to parallel prerequisites.
- Saved preferences and inference snapshots automatically compressed and capped at safe limits.
- Web terminals consolidated to single control-mode attach path, removing legacy PTY configuration.
August 2026
6 releases: 0.8.0 → 0.11.0
August brought major improvements to model routing, sandbox control, and session management, along with new harnesses like Devin and Grok Build. The month saw multiple refinements to web UI performance, collaboration features, and reliability.
- Sandbox dotfile hiding is now top-level only by default; untrusted-tree sandboxes must set cwdhiddenscanrecursive: true to maintain recursive masking.
- OpenTelemetry exporters and instrumentors now require the omnigent[tracing] extra instead of installing by default.
- Shared-session approval-authority and message-attribution features were reverted; session approvals are available to any shared editor.
Highlights
- Live model discovery now pulls from your active provider catalog and installed CLIs instead of fixed release lists.
- Smart Routing automatically selects the best model and harness for each task via Databricks AI Gateway.
- Run multiple sandbox providers at once with per-provider configuration and picker options.
- Manage sessions directly from the web UI: move between machines, fork sub-agents, and bulk operate on sessions.
- Native Claude Code sessions now support live permission mode switching and resume on their exact selected model.
New
- Live model discovery from active provider catalogs and installed CLIs (Kiro, Cursor, Pi).
- Smart Routing via Databricks AI Gateway for automatic task and model selection.
- Devin and Grok Build (xAI) as built-in harnesses.
- Detectloop and detectthrashing guardrail policies to catch stuck agents.
- Shared-session collaboration with delegated approval authority and author attribution.
- Nimble web builtins (nimbleextract and nimbleresearch) for structured data extraction.
- Blaxel sandbox provider and Kubernetes Jobs with automatic crash restart.
- Usage page tracking session costs, daily spend, and breakdowns by harness and model.
Improved
- Reworked web workspace with editor-style tabs, in-rail shells, full-screen toggle, and Mermaid diagram rendering.
- CLI groups and colorizes harness commands, with omni upgrade preserving extras and adding --extra/--target-version/--dry-run flags.
- Omnigent host survives dropped VPN and auto-reconnects on 401/403 errors.
- Conversation and terminal switching is now instant with background streams staying open.
- Session search is index-backed and bounded, no longer hanging or timing out.
- Parallel tool calls stream immediately instead of stalling for extended periods.
- Web and desktop hotkeys (Cmd/Ctrl+N for new session, ⌘K) work in terminal and composer.
- Claude Code sessions resume on exact selected model and honor /compact on sleeping runners.
Fixed
- Agent CLIs no longer inherit unrelated host credentials from the environment.
- Claude sessions routed through Databricks AI gateway now connect cleanly with proper headers.
- Approving a plan from the web UI now works for Claude Code.
- Background tasks no longer queue behind a Steer button.
- Stale Codex processes are reaped to improve reliability.
- Sessions and terminals ride out server restarts and network interruptions.
July 2026
5 releases: 0.4.0 → 0.7.0
July brought major platform expansion and UI polish to Omnigent, with iOS launch, desktop apps for Windows and Linux, and significant new capabilities for orchestration and automation. The month saw maturation across native harnesses and new ways to organize, navigate, and control agent workflows.
- omnigent[memory] renamed to omnigent[hindsight]; omnigent[memory] still works as backwards-compat alias but will be removed.
- Harness CLI path overrides standardized to OMNIGENT<NAME>PATH; legacy HARNESSPATH vars deprecated and slated for removal in v0.8.0.
Highlights
- Omnigent launched on iOS via the App Store, with Android coming soon.
- New Projects feature organizes sessions with shared settings; new Automations (Scheduled Tasks) page enables recurring agent tasks.
- Polly multi-agent orchestration now routes work across cursor, hermes, and opencode agents, with automatic intelligent model routing via server-side judge.
- Desktop app gained embedded browser for web project iteration, plus rich file viewing with PDF preview, find-in-file, and diff-viewer improvements.
- Slack integration lets you kick off and drive Omnigent sessions, including inline tool-call approval and Q&A.
New
- Generic ACP harness bridges any ACP-compatible agent to Omnigent's tools over an MCP relay.
- Harness plugin SDK lets you ship custom coding agents as installable Python packages without forking core.
- Message steering and queuing: line up, edit, reorder, and delete messages while a turn runs; queued messages auto-flush on idle.
- Voice dictation in composer with optional server-side transcription; works in Electron, Firefox, and Chromium with live streaming partials.
- Command palette (⌘/Ctrl+K) for jumping across sessions; session search shows highlighted matched-content previews.
- Git worktree support: start sessions in existing worktrees or generate fresh branches from the new-session composer.
- Kubernetes deployment with official omnigent-server-kubernetes image; runner Pods can schedule on arm64 via nodeselector.
- Slack integration for launching sessions and approving tool calls inline.
Improved
- All native harnesses (kiro, claude, qwen, GitHub Copilot) now support Omnigent tools, with kiro and claude reaching full parity including TUI approvals and chat cards.
- Intelligent model router now automatic, picking the best harness and model per turn via server-side judge reading live per-runner model availability.
- Appearance settings let you pick color themes (Omnigent, Dracula, GitHub, Catppuccin, Gruvbox), terminal theme, and independent UI and code editor/terminal font sizes.
- Theme editor now guided with shared accent/tint/contrast controls, sidebar translucency, new Nord palette, and randomize option.
- File viewing enriched with PDF inline preview and review comments, markdown and notebook previews, and persistent file panel.
- Large images no longer overflow context window or break resume/compaction; base64 attachment data stripped from stored and replayed history.
- Chat "Working…" indicator now clears reliably on idle, with gentle "send a message to wake it" hint for idle runners instead of error state.
- Sandboxed agents on Linux can now run tools managed by update-alternatives and reach hosts behind corporate MITM via system CA roots.
Fixed
- Forked native sessions (Claude Code, Codex, Pi, Qwen) now recover their prior history reliably.
- File and image attachments survive relaunches on remote host runners.
- On desktop builds without embedded browser, the Browser tab is hidden instead of showing a dead tab.
- Sessions resume reliably with proper history and attachment handling across contexts.
June 2026
2 releases: 0.2.0 → 0.3.0
June brought major expansion of agent harnesses and deployment options to Omnigent. The month saw seven new agent platforms added, broader sandbox support, and significant desktop and workspace improvements.
Highlights
- Support for seven new agent harnesses including Hermes, GitHub Copilot, OpenCode, Goose, Qwen Code, Kiro, and Kimi Code.
- Omnigent Desktop now automatically manages the server and runner.
- New Projects workspace lets you group sessions and organize your work more effectively.
- Multiple new deployment targets including Databricks Apps, AWS Bedrock, and Kubernetes.
- Native Windows support is now available for core features.
New
- Run agents on Google Antigravity with Gemini 3 Pro, Claude, and GPT-OSS models.
- Cursor agent support via first-party Python SDK with cost tracking.
- New sandbox providers: NVIDIA OpenShell, E2B, CoreWeave, and Podman.
- Deploy to Cloudflare Containers with D1 and R2, plus UBI9 image for RHEL/OpenShift.
- Projects workspace to group and organize sessions.
- Custom Agent Creation UI for building agents within the app.
- Native-harness parity features across all agents: compaction, cost/token tracking, resume, fork with history, model switching, and tool-approval cards.
- In-app MCP management from Agent Info panel.
Improved
- omni upgrade command for self-upgrading the CLI to the latest release with startup notices.
- Out-of-date CLI notices on startup.
- Omnigent Desktop automatically manages server and runner processes.
- Web UX improvements including shortcuts overlay, pinned-session hotkeys, and image lightbox.
- Per-harness run-mode memory in Projects workspace.
- Session sharing capability for agents.
- Native /compact command folded into omnigent setup for Codex and Pi login.
- Support for AWS Bedrock as a provider for agent execution.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Omnigent: pricing, features and alternatives · All changelogs