DeepSeek Harness Tests Claude Code Mods Compatibility, With Three Examples Running and Major Gaps Remaining

Three working examples, not full compatibility

DeepSeek released DeepSeek Harness v0.2.1-alpha.1 on October 3, 2026, adding an experimental bridge for extensions written for Anthropic's Claude Code. Three examples, Token Weather, Blast Radius and Replay Theater, are runnable through the bridge. Four other mods, diff, agents-md, sec-default and telemetry, are listed as unable to run because required events, interfaces or display capabilities are not supported.

That distinction defines the update. DeepSeek is testing whether its plugin architecture can express the capabilities exposed by Claude Code Mods, rather than promising that developers can move every extension between the two applications without changes. Mods must use the DeepSeek Harness loading mechanism, and the experimental bridge runs them without a sandbox.

DeepSeek's official release notes credit the compatibility layer to @tianyicui, the account of Harness team lead Cui Tianyi. The notes state the purpose of the experiment:

At this stage, the main goal is to verify that the Claude Code Mods API's capabilities are broadly a subset of what DeepSeek Harness plugins can do, rather than to offer users complete, practical compatibility.

The release also adds a more visible route for asking an agent to create plugins, optional tools for inspecting sessions, deployment improvements and fixes affecting plugin dependencies. Some changes require extension authors to update their code. Together, these additions make the alpha a test of extension portability and a maintenance release for an application increasingly organized around plugins.

A harness is the software surrounding a large language model that manages tools, sessions, files and task execution. The model supplies reasoning and generated text or code; the harness determines how that work is carried out.

How the bridge connects two extension systems

Claude Code Mods are JavaScript or TypeScript event handlers that run inside the coding application. They can respond when a user submits a prompt, when an agent calls a tool or when the application draws its interface. Depending on the exposed interface, a mod can observe an action, change it or take control of that step.

The DeepSeek bridge wraps a mod's registration function and mounts it as a Harness plugin. Its hooks are mapped to the Harness event system and lifecycle, the sequence through which components start, operate and stop. DeepSeek Harness uses Cordis, a framework for building software from services, dependencies and event subscriptions.

Through the bridge, supported mods can guard tool calls, rewrite prompts, add commands and tools, read session information and draw a band above the prompt box. These are changes to the agent application's behavior and interface, rather than simply instructions telling a model how to answer.

The word subset refers to the scope of those programming capabilities. DeepSeek is asking whether the operations offered to Claude Code mod authors can be represented within its broader plugin system. That is a narrower claim than proving identical behavior for every extension, and it does not establish that one application's model is more capable than another.

What the examples demonstrate, and what they do not

The three runnable examples exercise different parts of the bridge. Token Weather presents context usage as a weather forecast. Context is the information a model can use during a task, including instructions, conversation and tool results. The display gives users a visual indication of how much of that capacity has been consumed.

Blast Radius shows the likely effect of a risky command before execution, giving the user a chance to decide whether to proceed. Replay Theater lets users step through file modifications. Together, they test interface display, intervention before tool execution and access to the history of changes.

Those examples provide concrete evidence that useful extensions can be adapted across the applications. They do not establish complete support for the Claude Code Mods API. The four examples marked unable to run depend on capabilities that the bridge does not currently connect, including the /diff panel and agents.md parsing behavior.

There is also an installation boundary. A Claude Code mod cannot simply be copied into Harness and expected to work. It must be loaded through Harness configuration. Even an extension whose main hooks are supported may need changes to its registration, assumptions about session state or presentation.

The absence of a sandbox deserves particular attention. A sandbox is an isolation boundary intended to limit what running code can access or affect. Users should not interpret successful compatibility as a security certification, and the release does not promise that adapted mods execute in an isolated environment.

A rapid sequence of releases

The compatibility experiment arrived four days after the September 29 v0.2 preview, which introduced macOS and Windows desktop installers and a plugin management page. It followed Anthropic's introduction of Claude Code Mods at the start of October and was published during China's National Day holiday.

The sequence shows how quickly the project has moved from its initial developer preview to desktop distribution and extension experiments:

• March 2026: DeepSeek's Agent Harness team was formed, and Cui Tianyi joined DeepSeek.
• August 13: The v0.1 developer preview opened to developers worldwide, with source code released under the MIT license.
• September 29: The v0.2 preview introduced macOS and Windows desktop installers and a graphical plugin manager.
• October 3: Release v0.2.1-alpha.1 added the experimental Claude Code Mods bridge.
• October 4: Cui publicly explained the experiment's purpose, limitations and connection to the team's research goals.

The August launch attracted more than 70,000 GitHub stars in less than 24 hours. That measures attention to the repository, not active usage, successful deployments or plugin reliability. The later desktop release addressed a different issue: making the application and its extensions accessible without requiring every user to work through a command line.

Plugin creation becomes easier to reach

The new plugin management entry is named "Let Agent create a plugin" in the official release notes. It opens Creator mode while preserving the draft and begins work only after the user sends the request. That separation matters because entering a creation screen does not itself authorize the agent to start building.

Creator mode already allows an agent to write and install plugins in response to a user's requirements. The new entry makes that workflow easier to find within the management interface. The release also supports initial prompts that remain unsent in new sessions and preserves file, folder and session references when drafts are restored or moved between workspaces.

The September desktop preview had already introduced installation by npm package name. npm is a package registry commonly used for JavaScript software. Users could install, disable and uninstall plugins, as well as inspect their descriptions and origins, from the same page.

DeepSeek has said that about 60% of Harness users use third party plugins, measured among users of its official model API. That qualification matters: the figure describes that measured group, not necessarily everyone running the framework with any model provider. Expressed another way, approximately six in ten users in that group use external extensions.

The team plans to build an official plugin marketplace and improve extension discovery, acquisition and management. No launch date is given. The latest alpha improves package information in the meantime by showing a bundle's code source and current version more clearly.

Diagnostics and deployment receive practical fixes

An optional Developer Tools bundle adds raw session logs, navigation in both directions between chat groups and the conversation, and embedded Host diagnostics. The embedded interface currently uses English. These tools give developers more information for tracing agent behavior and investigating failures.

The release adds --public-url to specify the web address shown to users, opened in the browser and supplied to the model. It supports addresses behind a reverse proxy, including those with a path prefix. A reverse proxy routes incoming requests to an application that may run at a different internal address, so exposing the correct public address helps keep user navigation and model instructions consistent.

On Windows, the desktop application now lets the operating system assign its port by default. This is intended to avoid startup failures caused by reserved port ranges. The update also distinguishes a sign in request receiving no response from other authentication failures by showing a network check message.

Other fixes address queued messages after repeated stopping and resuming of a task, output and change displays for Bash, PowerShell and file operations, and multiline goal editing with Shift+Enter. Markdown previews now show YAML frontmatter as readable fields. Frontmatter is structured metadata placed at the beginning of a document, often used for properties such as a title or configuration.

These changes are less prominent than the compatibility experiment, but they address the everyday work of running agents: starting the application, preserving a request, inspecting what a command did and understanding why a task failed.

Extension authors face migration work

The alpha includes changes that can break existing extensions. Runtime invariant plugins and package ./invariant exports have been removed. Extensions and custom profiles using those diagnostic entry points must be updated.

Composer statistics are split into separate activity and usage extension entries. Plugins that replaced the previous statistics row need to update their registered IDs. Subpath plugins also no longer obtain display information from separate package.json files; text and icons must come from the corresponding subpath exports.

Dependency handling receives several fixes. Enabling a new bundle should no longer leave dependencies missing, while disabling or uninstalling bundles should no longer leave stale dependency mappings. Replacing an installed package version still requires a restart, an explicit limit on the current ability to change components while the application is running.

One interface fix prevents toggling a plugin from removing styles owned by other plugins until refresh. Automation tasks also become part of Web, with reminder tools available in Standard, Creator and PTC modes, but not Minimal mode or subagents. Existing tasks are preserved as retired experimental bundle selections are removed.

Why DeepSeek is pursuing an architecture of plugins

Cui says the principle that everything is a plugin was established before the project began writing code. In his description, models, tools, skills, sessions, sandboxes, file systems, the agent loop, task orchestration and the interface can be composed, replaced or extended.

That ambition goes beyond supporting an extension slot in a fixed application. It aims to let users and developers alter many parts of the system that carries out a model's work. Cui welcomed Claude Code Mods as a clever approach to extensibility after Claude Code lead Boris Cherny announced them, while distinguishing DeepSeek's broader architectural goal.

He identifies enterprise deployment as another use for plugins. Developers have built industry specific Harness plugins to create AI workbenches that connect business infrastructure and encode specialized workflows. Some of those plugins are not publicly released. Cui says the team supports that arrangement under the MIT licensed framework.

The distinction is between an open foundation and the extensions built on it. Businesses can use the shared framework while retaining proprietary integrations or industry knowledge in their plugins. The reports do not identify those deployments or provide figures for their scale or results.

Cui also argues that reproducing Claude Code or Codex would add little value because users can already connect DeepSeek models to those applications. His stated aim is to produce original work that influences how agent harnesses are designed, including innovations that major AI companies recognize and adopt.

Self modification is a research goal, not a finished capability

The longer term goal is for agents to change their own harness through plugins while running. Cui describes Creator mode as an early step toward that direction, rather than evidence that autonomous improvement has already been achieved.

He identifies two distinct areas of unfinished work. Engineering must improve the ability to add, remove and reload plugins during execution. Model research must teach agents that their harness can be modified and how to use those modifications to improve their work. Cordis supplies a foundation, but practical engineering problems remain.

The team is recruiting researchers with experience in large model research and training to pursue that work. Cui says a substantial part of his own time goes into recruiting and interviews. His background includes computer science at Zhejiang University, six ACM Asia regional competition gold medals and nine years at Jane Street before joining DeepSeek in March 2026.

No specific training method, measured gain from agent self modification or timetable for complete Mods compatibility has been disclosed. Cui says broader compatibility may be technically possible if the team decides it merits investment. That is a conditional assessment, not a commitment to a release date.

The immediate result is therefore concrete but limited: three runnable examples show that some Claude Code extensions can be represented in Harness, while unsupported mods, loading changes and migration requirements define the work still outstanding. The experiment tests an architectural idea; it does not yet deliver a universal extension ecosystem.

Key Points

• DeepSeek Harness v0.2.1-alpha.1 was released on October 3, 2026, with an experimental Claude Code Mods compatibility layer.
• Token Weather, Blast Radius and Replay Theater are runnable examples; diff, agents-md, sec-default and telemetry are not currently runnable.
• The experiment tests whether Mods API capabilities fit within Harness plugins, rather than providing complete practical compatibility.
• Mods require adaptation to Harness loading and run without a sandbox.
• The update adds a plugin creation entry, optional diagnostics, deployment improvements and changes requiring some extension migrations.
• About 60% of users measured through DeepSeek's official model API use third party plugins.
• Full compatibility, an official plugin marketplace and agent self modification research have no announced completion dates.