v0.7.2

Another patch primarily cleaning up support for Opus 4.5:

Model improvements:

• Opus 4.5 uses Anthropic's new effort parameter instead of thinking token budget
• Only show "1M Context" button for Sonnet 4.5

Also:

• Fixed bug where web/remote terminals weren't spawning in Electron
• Adds new web_fetch tool so Mux agents can read URLs

v0.7.1

This is primarily a patch release to expose Opus 4.5. Note that users can always try models before we "release" them by inputting the full model string via /model <model> or the model selector UI.

Also:

• Refined compaction prompt to focus on past, not future (0d923bf)
• Unified command handling between New Workspace creation and existing workspace chat pages (75c93e7)

v0.7.0

🪄 Features

Default model now set explicitly instead of defaulting to last used model:

Added light theme mode, toggled in top left:

Also:

• Integrated web terminal for SSH or browser-accessed workspaces (see coder/ghostty-web)
• "Tool: tool_name" AGENTS.md section for tool-specific instructions (docs)
• Support for Gemini 3.0 and Grok 4.1

🛠️ Fixes

• Optimized git status polling when you have dozens of workspaces
• Improved Anthropic caching
• Persisted costs/token usage across compaction cycles

v0.6.0

A major release with tons of goodies!

🪄 Features

Major: we have renamed the project from cmux to mux, but have taken great care in ensuring the migration is smooth for existing users. For example, we automatically symlink your old ~/.cmux directory to ~/.mux.

Tons of visual improvements on the UI, including revised agent sidebar and a more readable chat UI:

---

VS Code / Cursor extension for quickly jumping between IDE and Mux:

docs

---

File extension icons:

---

The New Workspace flow has been reworked so that the workspace name is generated automatically from first message:

---

Also:

• Added an experimental Windows build to the release artifacts
• Added support for Ollama and OpenRouter (docs)
• Users may define model-specific system instructions with Model: <model> in AGENTS.md (commit: 60b56ab, docs)
• Messages now queue until end of content block, leading to better model performance when interrupting frequently
  • This behavior seeks to mimic Claude Code to get the benefits of Anthropic's RL wrt interruptions
• Improved support for gpt-5.1-codex and the new mini model
• Updated the "codex" model abbreviation
  • More robustness to 500s/404s on strange Responses API behavior
• Optimized file_edit tool calls to reduce bad edits and malformed files

🛠️ Fixes

• b6fb712: retry barrier should no longer occasionally flash in before a stream starts
• e74e229: workspace cost now persists across compaction
• Fixed bug where file_edit diffs occasionally had multiple scroll bars
• Fixed many bugs on the auto-retry system, less spammy, fewer nonsensical auto-retries (e.g. does not auto-retry model not found)

🐞 Known Issues

• Nothing major to speak of!

v0.5.1

Couple of hotfixes:

a7a6b3c 🤖 fix: map baseUrl to baseURL for AI providers (#532)
70d85a5 🤖 fix: add mac zip artifact for auto updates (#515)

v0.5.0

This is a light release on the feature side due to our offsite last week.

🪄 Features

• General improvements to tool interface which has improved Terminal Bench scores ~5%
• Responsiveness improvements on small viewports, mobile development experience continues to get better
• 0f870a6: Code Review tab now defers syntax-highlighting until hunk is within viewport
  • multi-thousand line changes are now manageable in the Review tab

🛠️ Fixes

• d86727b: ‼️ startup freeze bug fixed
• 7a213e5: Fixed bug where bash commands would occasionally hang forever on SSH remotes

🐞 Known Issues

• #519 (Open Terminal is no-op in SSH and Web workspaces)

v0.4.1

Since v0.4.0

This is a patch release for calendar reasons, but contains a new feature that makes highly parallel workflow easier.

✨ Features

• Agents can now set statuses (with optional URLs for clicking)
  • It seems to pick decent emojis and URLs by default, but more in-depth customization is available via standard AGENTS.md guidance

• vim mode is now toggled with /vim, disabled by default for parity with Claude Code
• Code blocks now have line numbers and a copy to clipboard helper

🛠️ Fixes

• Fixed a bug where using "~" in the Project path would not resolve correctly and cause all workspace creates to fail
  • We both expand tildes by default now and only permit creating a project when the path exists and we can see it has a .git
• Fixed many bugs related to first usage when you don't have an API key set
  • User should now see clear instructions on how to resolve the issue vs. obscure errors
• Persistent chat errors (e.g. authentication, context_exceeded) are no longer auto-retried
  • Many general improvements to the retry and chat error behavior
• Interrupting a stream over SSH will now properly terminate in-progress bash tool calls and should complete nearly instantaneously
• Fixed a bug where editing a symlink would cause that symlink to resolve, most notably leading to issues when editing CLAUDE.md -> AGENTS.md type symlinks

⚠️ Known Issues

• Momentary UI freeze as tokenizer loads in on first UI launch - resolves within a few seconds
• Open Terminal doesn't work for SSH workspaces

v0.4.0

Since v0.3.0

✨ Features

• The headline feature of this release is SSH isolation for agents (docs)
• You can now run cmux as a stand-alone web app with: npm i -g @coder/cmux && cmux server
  • When paired with Tailscale or another secure tunnel this enables development from mobile devices such as phones and tablets

🛠️ Fixes

• Chat history is far more responsive on small viewports (e.g. when Review tab is open)
• An absolute ton of other bug fixes and performance improvements

⚠️ Known Issues

• Momentary UI freeze as tokenizer loads in on first UI launch - resolves within a few seconds

v0.3.0

Since v0.2.0

✨ Features

Integrated code-review system:

Maintained context display in thin viewports:

Real-time compaction streaming:

TODO lists:

Also:

• /compact now supports an optional -m <model> flag, e.g. /compact -m haiku to get a fast compaction
• /fork <name> to create a copy of the workspace, /new <name> to create a fresh one
• Syntax highlighting on all diffs
• Drag-and-drop image reading support for supported models
• Haiku 4.5 has first-class support via "haiku" abbreviation
• Rendering performance improvements

🛠️ Fixes

• Agents will no longer experience "amnesia" when resumed after an error
• Project Sidebar ordering persists through restarts
• Many compaction related bug-fixes

⚠️ Known Issues

• Startup can appear to freeze for up to 10 seconds, this goes away on second launch
• The experience is worth the freeze!

Telemetry

Note

This is the first version to ship with anonymized telemetry (details here)

v0.3.0-rc.2

Since v0.3.0-rc.1

✨ Features

• Context usage show up as thin vertical bar when viewport width is constrained
  • Context usage is now always visible when a chat is open
• /compact now supports an optional -m <model> flag, e.g. /compact -m haiku to get a fast compaction
• Haiku 4.5 has first-class support via "haiku" abbreviation
• Rendering performance improvements

🛠️ Fixes

• Project Sidebar ordering persists through restarts

⚠️ Known Issues

• Startup can appear to freeze for up to 10 seconds, this goes away on second launch
• The experience is worth the freeze!