Docker Just Fixed 90% of AI Coding By Releasing This

Highlights

• Docker introduced a "dynamic mode" for the MCP protocol to address scaling issues with AI agents managing numerous tools.

• The new dynamic mode significantly reduces token usage and speeds up AI agents by dynamically selecting and providing only necessary tool definitions.

• Docker's MCP catalog and gateway manage trusted MCP servers and autonomously provide tools, effectively solving context bloat.

• "Code Mode" allows AI agents to write custom JavaScript tools that can call other MCP tools, enhancing security and efficiency.

• Key benefits of Code Mode include secure sandboxed execution, improved token efficiency by abstracting multiple tool calls, and state persistence via volumes.

• Practical examples demonstrate creating custom tools like "multi search repos" and "GitHub to Notion" to streamline workflows and manage data efficiently.

Timeline

1. 00:00 - 00:42

   Problem with MCP Protocol Scaling

   The video begins by highlighting the evolution of the MCP protocol from handling a few local servers to hundreds with thousands of tools, creating a significant problem for AI agents. This scaling issue leads to inefficient token usage and agent performance degradation, a challenge first identified by Cloudflare and further detailed in a research paper by Claude. Docker has developed a solution, introducing a "dynamic mode" that promises to save tokens, accelerate agents, and enable new automation possibilities. This new approach aims to fix critical problems associated with the widespread use of MCPs in AI coding.

2. 00:42 - 01:22

   Docker's Solution: Addressing Key Problems

   Docker's solution addresses three core problems: identifying trustworthy MCP servers, preventing context windows from being filled with unused tool definitions, and enabling agents to efficiently discover, configure, and use tools autonomously. The speaker emphasizes the second problem, illustrating how an agent might only use 2-3 tools out of a thousand, yet all definitions bloat the context. Anthropic also previously discussed this issue, and Docker has now implemented a practical solution. This section sets the stage for understanding the specific mechanisms Docker employs.

3. 01:22 - 02:08

   MCP Catalog and Gateway for Trust and Management

   Docker's solution leverages its existing infrastructure, specifically the MCP catalog and MCP gateway. The MCP catalog lists verified and trusted MCP servers, making it easy for users to connect them within Docker. The MCP client (e.g., Claude code) connects only to Docker, which then centrally manages all connected MCP servers. This setup entirely solves the initial problem of trusting MCP servers by providing a curated and managed environment, ensuring that agents interact only with reliable tools.

4. 02:08 - 03:24

   Dynamic Tool Selection and Context Window Optimization

   To enable dynamic MCP usage, Docker implemented an MCP gateway with pre-built tools that autonomously utilize servers from the catalog. This means an agent connects to just one MCP, which then intelligently determines and brings only the necessary tool definitions into the context window, preventing bloat. New tools like MCP find, add, and remove allow agents to locate and manage servers dynamically. An example demonstrates searching GitHub repos, where the LLM initially returns excessive detail, highlighting the need for further context optimization.

5. 03:24 - 04:34

   Improving Token Efficiency with Dynamic Tool Provisioning

   This section delves into dynamic tool selection as a crucial aspect of token efficiency, referencing Claude's article on token consumption by tool definitions and intermediate results. The previous GitHub example showed how raw, detailed results unnecessarily fill the context window. Docker's MCP gateway addresses this by only providing the specific tools useful for a given session (e.g., just the 'search repo' tool instead of all 40 GitHub tools). This targeted provisioning significantly reduces the context window size, leading to substantial token savings and improved agent performance.

6. 04:34 - 05:15

   Introducing "Code Mode" for Advanced Agent Capabilities

   The video introduces "Code Mode," a groundbreaking feature implemented by Docker, based on Cloudflare's insights into the inefficiencies of traditional MCP usage. Code Mode allows AI agents to directly write and implement code using MCP tools, providing them with "code mode tools" that operate in novel ways. Essentially, it creates a JavaScript-enabled tool capable of calling other MCP tools. This innovative approach promises to unlock new possibilities for agent automation and efficiency, moving beyond static tool definitions.

7. 05:15 - 07:06

   Benefits of Code Mode: Security, Efficiency, and State Persistence

   Code Mode offers three significant benefits for AI agents. Firstly, it ensures complete security by running agent-generated code in a Docker sandbox, preventing any damage to the system. Secondly, it drastically improves token and tool efficiency; the model only needs to be aware of one new code mode tool, rather than all underlying tools, saving substantial context. Thirdly, it provides state persistence through volumes, allowing data to be saved between tool calls without flooding the model with large datasets, as exemplified by a data processing pipeline where only summaries or final results are sent to the model.

8. 07:06 - 08:14

   Practical Application: Multi-Search Repos Tool

   The speaker demonstrates Code Mode by creating a "multi search repos" tool. Initially, the agent combined multiple `find GitHub repo` calls into one, but the results were returned directly to the context window. To optimize this, the speaker instructed the agent to write all results to a text file, with the model only receiving a description of the repos, not the full details. This example showcases how Code Mode can streamline repetitive tasks and manage output efficiently, although the speaker notes the current limitation of manually saving and reusing such custom tools.

9. 08:14 - 09:09

   Practical Application: GitHub to Notion Workflow

   Building on the previous example, the speaker further demonstrates Code Mode by creating a "GitHub to Notion" tool. After connecting to the Notion MCP, the agent developed a tool to directly output GitHub search results into a Notion database, complete with fields for name, description, and date. This automation streamlines the process of organizing discovered repositories, ensuring that the model only receives minimal information (name and description) while the comprehensive data is stored externally. This highlights the power of chaining different MCPs to create complex, context-efficient workflows.

10. 09:09 - 09:55

    Getting Started and Conclusion

    The video concludes by encouraging viewers to explore Docker's MCP catalog to discover and chain together various MCPs for creating powerful workflows, emphasizing the benefits of saving tokens and preserving AI agent performance. Getting started is straightforward: users need to update their Docker version and ensure the Docker MCP toolkit is enabled in beta features. Once enabled, the catalog and new features are active by default, allowing users to connect their clients and immediately begin leveraging these advanced capabilities.