Jan v0.8.7 Released: Shell Tool Toggle and Subagents Added

At a Glance
| Item | Value |
|---|---|
| Repository | janhq/jan |
| Version | v0.8.7 |
| Published | 2026-10-20 |
| License | NOASSERTION |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
The latest version of Jan, v0.8.7, has been released. Jan is an open-source ChatGPT alternative that runs 100% offline on your own PC. Developed in Rust, it provides an easy-to-use desktop application environment for local LLMs.
The most impactful change in this update is the introduction of a “Shell" tool switch in the chat input area. This allows users to explicitly control whether to permit shell operations for the model, balancing security and convenience. For those who want to turn off the shell feature or restrict tool usage for specific tasks, this is an important update that allows for immediate switching decisions.
Breaking Changes and Deprecations
With this update, the hosted Exa MCP server, which had been added to the default list in the previous version (v0.8.3), is removed.
| Item | Old (v0.8.3 and later) | New (v0.8.7) |
|---|---|---|
| Hosted Exa MCP server | Included in the default server list | Removed during the initial startup cleanup |
This change affects users who are using Jan’s MCP features. However, servers manually added by users and Exa entries with custom API keys configured will be retained as they are. Because cleanup runs automatically upon the first launch, no special migration steps are required.
Key Changes
Shell Tool Toggle in the Chat Input Area A “Shell" switch has been added to the chat input menu. When this switch is turned off, the shell tool is no longer provided to the model. Meanwhile, Web Search, Fetch, and user-added MCP tools remain available. This feature appears when the shell can run within a sandbox and the model in use supports tool calling. It is set to on by default, but is useful when you want to prevent unintended command execution.
Attachment Image Size Limit Settings A new setting item has been added at Settings > Attachments > Max Image Size (MB). This allows users to specify the maximum size of attachable images within a range of 1MB to 100MB, with a default value of 10MB. Users can now flexibly adjust the limit when handling high-resolution images or, conversely, when aiming to reduce resource consumption.
Introduction of Built-in Subagents for Jan Agent Subagents with specific roles have been added within Jan Agent. Specifically, there are four types: explore (read-only by default), research, review, and debug. Individual models can be assigned to each of these, enabling advanced operations where optimal models are used according to the nature of the task. For engineers leveraging agent features, this is a major improvement that enhances workflow specialization.
Stabilization of Local Inference under Proxy and VPN Environments Chats using local models no longer fail in environments behind proxies or VPNs. Previously, system or environment variable proxy settings would capture requests even to Jan’s internal engine (127.0.0.1), causing errors. With this fix, communication to Jan’s own engine is now always made directly (no_proxy), ensuring stable operation regardless of network environments.
Forced Termination of Entire Turn via “Stop" Button Pressing the “Stop" button during chat execution now allows the entire turn to be terminated immediately, even during tool execution or while waiting for approval. This ensures users can regain control if the model continues outputting unnecessarily long responses or remains stuck waiting for tool responses. Additionally, adjustments were made so that the contents of web pages read by the built-in Fetch tool fit within the model’s context window.
Enhanced Approval for Command Execution on Windows In Windows environments, user approval is now always requested when command names contain ^ (e.g., cm^d). This is a security enhancement to prevent unintended execution via obfuscated commands. Furthermore, if a sandboxed shell command requests execution outside the sandbox, a prompt asking for user approval is now displayed each time.
Hardware Requirements
There are no explicit additions of newly supported model architectures or quantization formats mentioned in this version. However, important specification changes and limitations related to hardware environments and graphics drivers are included.
In environments equipped with NVIDIA GPUs, the graphics driver version is now automatically logged upon engine startup. This makes it easier to identify the cause if the inference engine crashes due to an outdated driver version, and displays an error message prompting a driver update. If issues occur in an environment performing inference using a GPU, please update your driver according to the message.
On the other hand, care is required regarding hardware compatibility in macOS environments. Macs equipped with Intel processors (Intel Macs) continue to lack support for running local models, carrying over from the previous version (v0.8.5). Please note that local model inference will not function if you are using an Intel Mac.
How to Get It
Jan is distributed as a desktop application, and updates can be performed by downloading and running the installer corresponding to your current environment. Existing user data and settings are carried over to the new version as-is when updating.
Note that upon the first launch, as mentioned earlier, a cleanup process to remove the hosted Exa MCP server automatically added in past versions (v0.8.3) will run automatically. MCP servers you added yourself and entries with individual API keys configured will not be affected and will remain intact.
Representative download methods and installer sources for each platform are as follows. Detailed and latest files are available from the official website (jan.ai) or GitHub Releases.
- Windows Environment
-
Official Installer: Download and run jan.exe. – Alternatively, you can download and update from the Microsoft Store.
-
macOS Environment
-
Universal Build: Download jan.dmg and place it in the Applications folder. – On macOS, the
mlx-serverprocess may remain in the background when force-quitting the app via SIGTERM or similar, so using the Cmd+Q shortcut is recommended when exiting. -
Linux Environment
- Debian / Ubuntu package: jan.deb – Package manager-independent executable: jan.AppImage – In addition, installation from Flathub and procedures for Arm64 are also provided. – As a note of caution, in standard Ubuntu 24.04 environments, sandbox functions are restricted due to the setting
kernel.apparmor_restrict_unprivileged_userns=1, and it has been reported that the “Shell" switch does not appear on the chat screen.
Additionally, if you are using the CLI tool, there is a known display bug where the version string is output as 0.8.5 when running the jan --version command, but the functionality itself operates normally as a v0.8.7 build.
Related Articles
- Jan v0.8.6 Released with Fixes for Linux, Turing GPUs, and MTP
- Jan v0.8.5 Released: Bundled Inference Engine and Key Changes
- Jan v0.8.8 Released: llama.cpp Upgrade and Security Enhancements
What to Read Next
- Follow this tool → Jan overview and release history (71 releases tracked)
- Other uis and desktop apps → Open WebUI / TextGen

