Claude Code Opens Mods with Highest Privileges; DeepSeek's Cui Tianyi Says It Converges with DSH but Scope Differs

On October 1, Anthropic officially announced the introduction of the Mods (Modifications) mechanism for Claude Code, allowing developers to deeply customize the agent's interface and runtime logic through in-process function hooks. Users don't even need to program themselves—they can have Claude Code write and install Mods for itself and hot-load them in the current session.

This update quickly drew responses from peers. Cui Tianyi, team lead of DeepSeek's agent foundation DeepSeek Harness, reposted and commented on social media that Claude Mods and DeepSeek Harness are 'converging at the underlying level—all roads lead to the same destination'—but the scope of what the two open up is entirely different. It should be noted that this is Cui Tianyi's personal opinion, not an official DeepSeek statement, and no response from Anthropic appears in the sources.

Cui Tianyi's background adds weight to this comment: he was formerly a technology expert at top quantitative fund Jane Street, author of the classic open-source tutorial 'Nine Lectures on the Knapsack Problem,' and joined DeepSeek this spring to lead Harness. His judgment is: Mods in Claude Code are essentially still local plugins patching an established software architecture, and Claude Code currently opens only a subset of functionality as Mods; whereas DeepSeek Harness, from its very first day, built its entire runtime on an 'Everything is a Plugin' architecture.

Thus, a product update evolved into a route divergence over 'to whom runtime control is opened, and to what extent.'

Before this update, Claude Code's extension mechanisms mainly relied on three relatively fragmented channels: MCP (Model Context Protocol) exposes external read-only or execution tools via standard JSON-RPC; Skills are purely prompt-driven operation guides; and Settings Hooks launch independent external subprocesses to run shell scripts at specific trigger points. The pain point of this approach is the inability to intrude into the main process—external scripts can neither obtain and modify runtime intermediate states in real time nor perform pixel-level reconstruction of the terminal presentation interface.

The core change of Mods lies in evolving the extension mechanism from 'out-of-process calls' to 'in-process interception.' Technically, a Mod is a piece of code written in JavaScript or TypeScript that resides directly in Claude Code's main process, intervening in the execution flow via event hooks.

In terms of capabilities, Mods have rendering rights over the terminal interface: developers can draw a code change diff panel beside the main conversation area, render a multi-agent call tree, display token consumption rates in real time in the status bar, and even redraw the system's default execution waiting animation. They can also perform bidirectional interception of tool calls—intercepting requests before a tool triggers and popping up a confirmation mechanism for the user, and automatically filtering and desensitizing sensitive credentials (such as API Keys) after execution completes but before results are fed back to the large model. Additionally, Mods support registering custom /command commands that return locally in seconds without invoking the large model, and can route requests to cheaper or faster models based on task complexity within the execution loop.

One noteworthy detail: Anthropic's recently launched /diff panel and agents.md support for multi-agent collaboration have both been refactored at the underlying level into built-in Mods. This means Anthropic itself is also using the Mods mechanism to decouple product functionality—though this claim is paraphrased in the source and is not directly confirmed by Anthropic's official documentation.

Anthropic also provided three intuitive examples: Token Weather draws a weather-forecast-like graphic above the input box, dynamically showing current context window usage; Blast Radius pauses operations when Claude Code is about to execute potentially impactful commands like rm -rf or force push, showing the user what might change and offering continue or cancel buttons; Replay Theater provides a /replay command, allowing users to step through what modifications Claude made to files in the previous round of operations.

The core logic behind Cui Tianyi's 'all roads lead to the same destination' judgment is that both Anthropic and DeepSeek have realized that relying purely on 'large model + static prompts' cannot solve stability problems in complex engineering scenarios. To enable agents to deliver reliably in real codebases and terminal environments, runtime control must be opened up to developers.

But the two teams have given different-dimensional solutions on 'to whom to open up, and to what extent.' According to Cui Tianyi's description, Claude Code remains a productization-first black-box tool; its core execution mechanisms, session management, and main state machine do not allow external modification. What Mods open up are only a few officially preset facets (interface rendering, tool filtering, prompt injection).

In contrast, DeepSeek Harness is built on the open-source microkernel framework Cordis, with no privileged core in its architecture—every system component is implemented as an independent plugin: basic capability plugins include model adapters (supporting DeepSeek, open-source model endpoints, and other commercial models), tool sets, external Skills, execution sandboxes, and local persistent storage; process and system plugins include session event logs, task scheduling systems, and the execution step loop (Agent Loop); among presentation-layer plugins, the frontend UI (whether web interface or terminal interaction) is itself merely a pluggable Bundle.

Based on this, Cui Tianyi joked: the discipline that used to study how to attach tools and prompts to models externally was called 'Plugin Engineering,' and with the comprehensive devolution of agent control, this technology might now be renamed 'Mod Engineering.'

It should be pointed out that the above descriptions of DeepSeek Harness's architecture all come from Cui Tianyi's personal statements; the source provides no third-party verification, nor any data on developer adoption rates, performance, or ecosystem scale for either product.

DeepSeek Harness's four official operating modes are Standard (full coding), Code (programming interface), Minimal (minimal evaluation), and Creator (creator mode), among which the most striking is precisely the Creator Mode mentioned in Cui Tianyi's tweet.

According to the source description, Creator Mode grants the agent the introspective modification capability to 'write its own plugins on the spot': in terms of runtime introspection, the Agent has read access to the current system runtime environment and can inspect which model adapters it has loaded, which tools are mounted, and what scheduling strategy is enabled; in terms of in-memory on-the-spot coding, when facing complex tasks beyond its current capabilities, the Agent can directly write a piece of Cordis plugin code in memory; in terms of hot-mounting and persistence, once the new plugin is written, the system can immediately complete mounting, activation, and verification within the sandbox, and once verified as viable, the Agent can solidify it in real time into a brand-new preset mode.

Through this approach, one can modify the UI, adjust the Agent Loop's behavior and operating mechanisms, add new tools, or introduce new task orchestration methods. Using the metaphor from the source: Claude Code's Mods are 'functional peripherals' that human engineers assemble for the agent, while DeepSeek Harness's Creator Mode gives the agent the possibility of 'rewriting its own organs and operating logic on the spot.'

The boundaries of this comparison also need to be marked: the actual capabilities and stability of Creator Mode lack third-party verification, and the source does not disclose the complete API and permission boundaries of Claude Code Mods' function hooks. The actual differences between the two in the degree of runtime control openness lack quantifiable comparison.

This technical confrontation did not come out of nowhere. DeepSeek Harness was initiated in May, open-sourced its developer preview v0.1 in August, and the developer preview was officially open-sourced in mid-August. Since Cui Tianyi joined DeepSeek this spring and formed a team, DeepSeek began to fill in the missing link of 'transforming a smart large model into real execution capability.'

The source mentions a formula circulating internally at DeepSeek: Model + Harness = Agent. If the large language model is compared to an engine, Harness is the chassis, transmission, steering wheel, and brake system. Over the past year, global large model vendors have mainly competed on model parameters, inference costs, and context length; the launch of Claude Mods and the open-source advancement of DeepSeek Harness signal that the competitive dimension is extending to the runtime layer.

According to the source's summary, Anthropic, relying on its mature product form, attempts to lower the customization threshold for professional engineering teams through Mods, shaping Claude Code into an extensible productivity benchmark; DeepSeek, on the other hand, attempts to build a boundaryless open-source agent foundation through a thoroughly microkernel plugin-based architecture.

It should be noted that related information mentioned in the source, such as the launch of the DeepSeek Harness desktop preview and V4-Pro programming capabilities, has unclear details and will not be elaborated in this article.