- 13 min read
- 8 key claims
- 5 sources cited
- For Developers, AI makers, Businesses
Overview
Open WebUI's v0.13.x releases reorganize the project around a persistent workbench rather than a single chat window. v0.13.0 supplies the workspace itself: files, multiple terminal sessions, artifacts, previews, voice calls and comments open as tabs, with tabs that can be added, reordered and used to create files in a selected terminal [1]. A Desktop tab exposes a connected Open Terminal's desktop for mouse, keyboard, clipboard and audio control, and a Browser tab provides navigation, multiple tabs and a mobile viewport, with Chrome mode giving the user and a model a shared browser session for automation [1].
The workspace also carries development and review tools. A Git panel can review diffs, stage, discard and commit changes, manage branches and worktrees, push and pull, browse history and work with GitHub pull requests when the selected terminal's folder is a Git repository [1]. Side chats, invoked with "/side" or "/btw", open a temporary conversation using the existing chat as context and disappear when their tab is closed [1]. Administrators can enable comments with replies, reactions and mentions across chats, notes, models, knowledge bases, prompts, skills and tools, and a notification inbox collects comment activity, mentions, failed automation runs and upcoming calendar events [1]. MCP tools that include an interactive app can render that app in the chat inside a sandboxed frame limited to sites the server declares, with user confirmation required for each requested tool call [1].
v0.13.1 changes what can appear in the model picker. Coding agents such as Claude Code, Codex and Pi appear on Open Terminal connections that offer agents, with reasoning-effort and service-tier selection and in-chat approval of tool requests [2]. The same release introduces Tether, a desktop companion app for macOS, Windows and Linux that connects a user's computer to Open WebUI and initiates the outbound connection, so no ports need to be opened; the project states it resumes terminal output after a dropped connection and can deliver desktop notifications when no Open WebUI page is open [2]. Tether signs in with a renewable session that expires after 90 days of inactivity by default, configurable through "SESSION_EXPIRES_IN" [2].
Tool governance and file handling change alongside. An Auto tool-permission mode sits beside Full access and Ask for approval; it is off by default and uses a review model that sends the tool, arguments and recent messages to a configured scoring API, falling back to approval for calls it cannot review [2]. Files with extracted text below a default threshold of 50,000 characters are included in full rather than through retrieval, both when preloaded and when read via the file-search tool, and the threshold is configurable through "RAG_FULL_TEXT_CHAR_THRESHOLD" or admin Documents settings [2]. On large screens, workspace tabs can split into resizable side-by-side panes, with each chat's tabs, split layout and selected terminal saved privately per user and restored on another device [2].
v0.13.2 fixes opening Open WebUI over plain HTTP at a LAN or Docker address, which previously remained on a spinner with a "crypto.randomUUID is not a function" error, and fixes creating, replying to and editing comments and editing Markdown files containing rendered blocks over plain HTTP [5]. It also makes the Allow Side Chat switch visible even when Temporary Chat is off so administrators can configure the two permissions independently [5].
v0.13.3 adds coding agents to automations: the automation editor can select a terminal, working folder and coding agent, and scheduled runs verify that the terminal and agent remain available before starting [4]. The terminal and working folder can also change after messages are sent, including in temporary chats, with both controls disabled while a reply is generating [4]. Brazilian Portuguese translations are added [4]. The release also stops connecting a terminal from adding its coding agents to admin models, base-model selectors, Arena candidates, task-model settings, group defaults, sub-agent settings, channel mentions or the Playground, and agents cannot be saved or imported as custom models or their base models [4].
v0.13.4 adds Tether as the replacement for Direct Connections, with models from Ollama, LM Studio or any OpenAI-compatible server, plus OpenAPI and MCP tool servers and Open Terminal servers, appearing in the model selector, Tools list and terminal menu for the signed-in user across devices; the project states that addresses, commands and keys never leave the user's computer [3]. With Terminals 0.4.0 or later, Open WebUI can provide sandboxed computers on demand, configurable as shared or per-user, per-chat or per-model instances, with image, CPU, memory, storage, network profile, environment variables, start directory and setup script settings [3]. The same release expands the coding-agent roster to Hermes, Antigravity, GitHub Copilot, Goose and Qwen Code, alongside Codex, Claude Code, Cursor, Grok, OpenCode, Cline, Gemini CLI and Pi [3]. Agent replies can now be regenerated and earlier messages edited, and forked agent chats continue for Codex and Claude Code while other agents restart from the chat so far, retaining file changes already made [3]. Notes, channels, the workspace, automations, the calendar, the playground, the home page and new chats from pinned models can open as tabs beside the current chat [3]. A message sent while a reply is running is queued on the server, kept if the page is left, and answered at the reply's next step with work completed so far [3].
Why it matters
For developers, the central change is that a terminal, its files, version-control review and a shared browser session sit in one workspace rather than in separate tools [1]. A coding agent can run against a repository while the Git panel shows diffs, branches, history and pull-request work for the same folder [1].
For makers and small teams, terminals, scheduled automations and comments on chats, prompts and knowledge bases make Open WebUI usable as a shared project surface [1][4]. Comments support replies, reactions, mentions and anchoring to a selected chat passage, and the notification inbox gathers comment activity, mentions and failed automation runs [1].
Local operation is the other substantive shift. Tether makes models from Ollama, LM Studio or any OpenAI-compatible server, plus OpenAPI, MCP and Open Terminal servers, reachable from any browser the user signs into; the project states that addresses, commands and keys never leave the computer [3]. With Terminals 0.4.0 or later, sandboxed computers can be handed out on demand without an administrator setting up each one [3].
Tool governance is becoming an administrative concern rather than a per-call decision. Auto mode reviews tool calls and falls back to human approval when it cannot review a request [2], while Auto for Codex, Claude Code and Grok lets those agents approve routine actions and escalate selected cases [3]. Administrators can set default tools for models and gate comments, side chats, notifications and sandbox networking [1][2][3].
For self-hosted deployments, v0.13.2 addresses a basic access problem: plain HTTP over a LAN or Docker address now loads chats and supports comments and Markdown editing [5]. The project also states that large chats open about twice as fast, that read or unread marking and tag removal are 4 to 6 times faster, and that typing after a 16,000-line paste improved from 8.6 seconds to 74 milliseconds per keystroke [2]. Those figures come from the project's own release notes rather than independent testing.
Opportunities
A developer could run a coding agent against a real repository, review diffs in the Git panel, push a branch and open a pull request without leaving the chat workspace [1]. A small business could give each client or project its own on-demand sandboxed terminal, configured with a setup script that clones a repository and installs dependencies, and share terminal instances across users and groups [3].
Tool builders could package an MCP server with an interactive app and surface the app directly in the chat, with the server restricted to the sites it declares and each requested tool call requiring user confirmation [1]. An MCP server distributed as a local command over stdio can be launched through Tether without a proxy such as mcpo or an open port [3].
Coding agents in the model picker allow an agent to be selected for a conversation, given a reasoning effort and service tier, and have its tool requests approved in the chat [2]. By v0.13.3, agents can also be selected inside an automation with a saved terminal and working folder, and scheduled runs check that both remain available before starting [4].
The tab system could support a review workflow: notes, channels, automations, the calendar, the playground and new chats open as tabs beside the current chat, and a user's tab layout is saved per chat and restored on another device [2][3]. For audiences outside engineering, comments, the notification inbox and side chats offer a lighter-weight review and branching mechanism for prompts, knowledge bases and notes [1].
Risks
Agent execution on a personal machine is the main risk surface. Tether's sandbox is stated to restrict terminal-launched processes to one chosen folder with network access off by default, but local MCP commands run with the user's own account access outside that workspace sandbox [3]. A compromised or poorly configured tool server therefore sits outside the folder boundary. If sandboxing cannot be enforced, the terminal stays stopped and browser and desktop control remain off [3]; that is a fail-closed design, but its effectiveness on each platform is not independently documented.
Auto tool permissions introduce a dependency on a review model. The review sends the tool, its arguments and recent messages to a configured scoring API and runs a call only when it judges no confirmation is needed; errors, timeouts or excessive context fall back to approval [2]. This can reduce interruptions, but a misconfigured or manipulated review model could approve actions a user would not have approved manually. The release notes do not describe audit trails, accuracy rates, or how the review model is isolated from the agent being reviewed.
Adoption and upgrade costs are substantial and uneven, given the version floors for Tether, Terminals and individual Open Terminal agent features [3]. Tether replaces Direct Connections, and the prior status and migration path of Direct Connections are not described [3].
Behavioral change can disrupt existing workflows. In v0.13.1, agent chats could not be regenerated, continued, or edited and resent [2]; by v0.13.4, replies can be regenerated and earlier messages edited, and forked agent chats continue for Codex and Claude Code while other agents restart from the chat so far [3]. The tool-call message layout changed so custom pipes see hidden tool results rather than the prior arrangement [2]. Connecting a terminal stopped adding its coding agents to admin models, base-model selectors, Arena candidates, task-model settings, group defaults, channel mentions and the Playground [4]. Administrators who relied on those surfaces would need to rework their configuration.
Cost effects are directionally favorable but unquantified. The project states that tool chats cost less because citation instructions, memory recall, skills lists and attached knowledge are no longer rewritten into every tool round, and that batched Tool Search calls save model turns [2]. No price figures or measured savings are given. Small files below a 50,000-character threshold are now included in full rather than through retrieval, which trades retrieval complexity for larger prompts on long documents [2]. Local MCP commands run under the user's account, so permission errors and side effects on local files remain a possibility even though the project says local commands run outside the workspace sandbox [3].
Context
The v0.13.x line builds on two companion systems that the releases name explicitly. Open Terminal is the terminal-side component: desktop control, browser support and agents all depend on features it provides [1]. Terminals is a separate system for managed terminal instances, requiring version 0.4.0 or later for on-demand sandboxed computers and 0.4.1 or later for start and stop [3].
A second layer of version floors shapes what users can actually switch on. Tether support requires version 0.1.14 or later, or 0.1.16 for MCP servers, tool servers and terminals with "ENABLE_TOOL_SERVERS" enabled [3]. Sandboxing on a personal computer needs Tether 0.1.13 or later, while local MCP servers over stdio need Tether 0.1.16 or later [3]. Individual agent features need specific Open Terminal builds: 0.17.19 or later for the new agents and Auto for Claude Code and Grok, 0.17.9 or later for reasoning effort on Claude Code, Pi and ACP agents, and a later release than 0.17.20 for Claude Code continuation [3]. Start directory and setup script options need Open Terminal 0.17.7 or later in the terminal image, and network profiles are administrator-defined through "TERMINALS_NETWORK_PROFILES" [3].
The releases also record changes in behavior across the line. In v0.13.1, agent chats could not be regenerated, continued, or edited and resent [2]. By v0.13.4, replies can be regenerated and earlier messages edited, and forked agent chats continue for Codex and Claude Code while other agents restart from the chat so far [3]. The custom-pipe message layout changed because tool-using chats no longer rewrite citation instructions, memory recall, skills lists and attached knowledge into the system prompt or message on every tool round [2].
One point is unresolved across the notes. v0.13.0 describes side chats as temporary conversations that use the chat so far as context and disappear when their tab is closed [1], while v0.13.4 describes side chats as running the same chat as the main view with shared message controls, sources, token usage, follow-up suggestions, slash commands, queued messages, chat variables and web search settings [3]. The later release does not restate whether a side chat still disappears on close.
Known, claimed, inferred
Consolidation into a workbench. The tabbed workspace, per-chat saved layouts and split panes suggest a move away from a single-conversation model toward a persistent workbench spanning chat, code, browser and desktop [1][2][3]. This is a reasoned inference from the interface changes, not a stated project goal.
Agents as session-scoped models. The decision to confine terminal agents from global model-selection surfaces suggests the project treats agent availability as tied to the connected terminal and chat rather than as a globally registered model [4]. Coding agents appear in the chat model picker on connections that offer them [2], and their roster expanded in v0.13.4 [3].
Rapid change in agent semantics. The progression from agent chats that could not be regenerated or continued [2] to regenerated and forked agent chats [3] shows that agent behavior changed materially within the line, so upgrade planning should account for behavioral changes rather than only new features.
Governance shifted toward administration. Auto permissions, default model tools, the independent Allow Side Chat switch and administrator-defined network profiles place more control in admin settings than in individual chat actions [2][3][5]. The review-model dependency in Auto mode is the new component in that chain.
Local reach widened, with boundaries drawn unevenly. Tether extends the set of reachable local models and servers while the project states that addresses, commands and keys never leave the computer [3]. The sandbox covers terminal-launched processes on macOS and supported Windows devices, but local MCP commands run outside that sandbox under the user's own account [3].
Limitations & uncertainties
The available material covers release notes for v0.13.0 through v0.13.4 only. No release dates, adoption figures, user counts or third-party test results are given. The v0.13.0, v0.13.1 and v0.13.4 entries are truncated, so their complete feature lists are unknown [1][2][3].
No independent benchmarks exist for the stated speed, cost and keystroke improvements, and no security audits, known vulnerability reports, regression rates, upgrade breakage reports or migration requirements between these versions are documented. Licensing and pricing for the coding-agent, MCP and terminal features are not stated. Prior status and migration handling for Direct Connections, which Tether replaces, are unknown [3].
Platform coverage of the Tether sandbox is qualified as macOS and supported Windows devices, with no corresponding detail for Linux [3]. The evidence does not describe the overall significance or impact of the v0.13.x line relative to other Open WebUI releases. Whether side chats still disappear on close after the v0.13.4 change is also unstated [1][3].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementOpen WebUI v0.13.0 introduced a tabbed workspace for files, terminal sessions, artifacts, previews, voice calls and comments, with tabs that can be added and reordered.[1]
- Primary-source or vendor statementThe workspace includes a Git panel for diffs, staging, commits, branches, push and pull, history and GitHub pull requests, plus a Desktop tab and a Browser tab with shared browser sessions in Chrome mode.[1]
- Primary-source or vendor statementFrom v0.13.1, coding agents such as Claude Code, Codex and Pi appear in the model picker on agent-capable Open Terminal connections, with reasoning-effort and service-tier selection and in-chat approval of tool requests.[2]
- Primary-source or vendor statementv0.13.4 adds Tether as the replacement for Direct Connections, exposing local models, OpenAPI and MCP tool servers and Open Terminal servers for the signed-in user across devices; the project states addresses, commands and keys never leave the user's computer.[3]
- Primary-source or vendor statementThe Auto tool-permission mode is off by default, sends the tool, arguments and recent messages to a configured scoring API for review, and falls back to approval for calls it cannot review.[2]
- Primary-source or vendor statementThe project states that large chats open about twice as fast, read or unread marking and tag removal are 4 to 6 times faster, and typing after a 16,000-line paste improved from 8.6 seconds to 74 milliseconds per keystroke; these figures come from the release notes rather than independent testing.[2]
- Primary-source or vendor statementv0.13.3 adds coding agents to automations with a saved terminal and working folder, and scheduled runs check that the terminal and agent are still available before starting.[4]
- Primary-source or vendor statementLocal MCP commands launched through Tether run with the user's own account access, outside Tether's workspace sandbox.[3]
Sources
- 1open-webui/open-webui v0.13.0●github.comretrieved Nov 18, 2026
- 2open-webui/open-webui v0.13.1●github.comretrieved Nov 18, 2026
- 3open-webui/open-webui v0.13.4●github.comretrieved Nov 18, 2026
- 4open-webui/open-webui v0.13.3●github.comretrieved Nov 18, 2026
- 5open-webui/open-webui v0.13.2●github.comretrieved Nov 18, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.