Kimi API Now Natively Compatible with Codex and Claude Code

On September 2, Moonshot AI announced that the Kimi API has officially begun supporting the OpenAI Responses API format and the Anthropic Messages API format. Users of Codex and Claude Code can now connect directly to the Kimi model lineup via a custom model provider without going through format conversion or a local proxy. The move is attracting attention as an interoperability initiative that expands options for coding agents.

According to reporting by TMTPost’s TMTPost Official Account, the aim of this update was described as follows:

The Kimi API now officially supports the OpenAI Responses API format and also supports the Anthropic Messages API format. Users of Codex or Claude Code can connect directly via a custom model provider without going through format conversion or a local proxy

Eligible models include kimi-k3, kimi-k2.7-code-highspeed, kimi-k2.7-code, kimi-k2.6, and others. Developers can call Kimi-family models simply by modifying their existing agent configuration.

Overview of the Announcement and Eligible Models

This announcement responds to requests from developers who wanted to use the Kimi API with Codex and Claude Code. On September 2, Moonshot AI simultaneously released compatibility for two major API specifications.

One is support for the OpenAI Responses API format. The other is support for the Anthropic Messages API format. Both are interfaces widely adopted by agentic coding tools.

Models available for connection include kimi-k3, kimi-k2.7-code-highspeed, kimi-k2.7-code, and kimi-k2.6. From flagship models to high-speed models specialized for code generation, users can choose according to their needs.

Previously, using models from different providers with Codex or Claude Code required developers to prepare their own request-format conversion layer or a local proxy server. With this native support, that workaround is no longer necessary.

Moonshot AI’s approach represents the model provider adapting to the agent-side interface. For developers, this lowers the barrier to adoption.

Technical Compatibility Details

The Responses API is an input/output format adopted by OpenAI for its agent products, including Codex. It is characterized by a design that handles tool calls and sequential reasoning processes in a structured way.

The Messages API is a conversational API used by Anthropic for Claude Code. It manages system prompts, user messages, and tool execution results with clear separation.

By natively interpreting both formats, the Kimi API allows clients to use it simply by swapping the endpoint and authentication credentials. The expected operation is to enter the Kimi API information in the custom model provider field.

Eliminating the format conversion layer also helps reduce latency and suppress errors. It avoids parameter omissions and type mismatches that tend to occur in conversion layers.

In addition, since there is no longer a need to keep a local proxy running, the complexity of the development environment is reduced. This also simplifies configuration when running in CI environments or inside containers.

For models specialized for code generation, such as kimi-k2.7-code-highspeed, compatibility with the agent’s iterative edit-and-execute loop is critical. Native support will encourage adoption in workflows that assume such rapid trial and error.

Impact on Developer Experience

Both Codex and Claude Code are coding agents that run in a local environment and autonomously perform file editing and command execution. They work closely with terminals and editors to handle multi-step development tasks.

An increase in the models available for these agents means more choice for developers. They will be able to compare models and choose based on suitability for specific languages or frameworks, response speed, cost, and other factors.

The Kimi series in particular has been recognized for its long-context processing and code generation in both Chinese and English. Its use is anticipated in scenarios involving domestic development environments and documentation.

Direct connection via a custom provider enables an operation where models can be swapped while maintaining existing workflows. There is no need to rebuild prompts or tool definitions.

On the other hand, when switching models, it is necessary to verify output stability and tool-call accuracy. Since agent behavior depends heavily on the underlying model, evaluation during migration is essential.

This update illustrates a structure in which the agent-side ecosystem promotes competition among model providers. The standardization of interfaces is creating an environment that more directly tests model differentiation.

Related Developments Announced Simultaneously

On the same day, September 2, a series of agent-related announcements followed in China. Tencent’s WorkBuddy officially launched its open platform open.WorkBuddy.cn and, together with more than 100 ecosystem partners, opened up Agent capabilities to smart hardware and industry-specific applications. Plaud, Rokid, Insta360, iFlytek, Anker, Mammut, JD.com’s Jingzao, and others were among the initial partners, with more than 30 applications connected across more than 20 fields including finance, law, healthcare, and education.

Doubao Work also announced new features on the same day supporting parallel execution of multiple Agents and computer operation on Mac. These include parallel processing where a supervising main Agent decomposes tasks and distributes them to specialized sub-Agents, and a feature that performs GUI operations through screen understanding even in environments without MCP, API, plugins, or CLI.

These moves indicate a trend of expanding agents from a single chat interface to execution environments that include hardware and OS operations. Kimi API’s interoperability support can also be positioned as part of building the infrastructure for agents to operate on the premise of diverse models.

Overseas, CrowdStrike announced an expanded collaboration with OpenAI, extending enterprise security protection to the Codex agent and introducing GPT-5.6 Cyber to the Falcon platform. This is a case that shows security integration has become an indispensable element in the enterprise adoption of coding agents.

Background to Growing Interoperability

Large language model APIs initially each presented proprietary specifications. However, with the spread of agents and development tools, OpenAI-compatible and Anthropic-compatible interfaces have begun to function as de facto standards.

For model providers, supporting the API formats adopted by major agents is a shortcut to expanding adoption. This is because developers can try new models while maintaining their existing toolchain.

Moonshot AI’s latest update follows this trend. By supporting both of the two major formats — the Responses API and the Messages API — it opens its doors to developers on both the Codex and Claude Code sides.

This progress in interoperability makes model performance competition more transparent. Since different models can be compared on the same agent and the same task, the reproducibility of benchmarks is improved.

At the same time, the division of responsibility for who maintains the API compatibility layer is also becoming clearer. Previously, there were many cases of reliance on unofficial community-built conversion adapters, but official support improves maintainability and reliability.

Going forward, similar native support may spread to other model providers. Once agent-side interfaces stabilize, support on the model side is expected to become a prerequisite rather than a differentiator.

Future Outlook and Points to Note

The native support for the Kimi API has the effect of reducing technical friction at the time of introduction for developers considering the use of coding agents. In particular, teams already operating Codex or Claude Code can begin evaluation without preparing additional infrastructure.

However, API format compatibility and actual agent performance compatibility are separate issues. Success rates for tool calls and instruction-following ability in long contexts depend on model-specific characteristics. Even after official support, verification on individual tasks will still be necessary.

In addition, running an agent involves not only the model’s inference cost but also token consumption associated with tool execution and iterative loops. Cost management and rate limit design will be important considerations at the operational stage.

In terms of security, due to the nature of agents accessing local files and command execution, it is necessary to understand the scope of data sent to the model provider. In enterprise environments, it is required to confirm data retention policies and log handling.

The model lineup presented by Moonshot AI consists of general-purpose models, code-specialized models, and high-speed variants to be used selectively. Operation that selects a model according to the agent’s purpose is expected to become the standard going forward.

Editorial Opinion

In the short term, we expect trial use of Kimi models to accelerate among development teams that have already adopted Codex or Claude Code. This is because evaluation costs will drop significantly as format conversion and proxies are no longer required. Within the next 3 to 6 months, comparative verification on routine tasks such as code completion, refactoring, and test generation may become more active. We assess that the role of unofficial community adapters will diminish, and benchmark sharing based on official compatibility will advance.

In the long term, we see API interoperability increasing the fluidity of model selection and promoting the separation of agents and models. Over a span of 1 to 3 years, horizontal specialization may strengthen, where agents compete on interfaces and models compete on performance and cost, rather than vertical integration where a specific agent is tied to a specific model. While developers will have more choices, we believe that developing evaluation methods and governance to absorb behavioral differences between models will become a new challenge.

The remaining question is how complete the compatibility support is. Which versions and which parameters of the Responses API and Messages API are faithfully reproduced remains to be seen.

References

• “【钛晨报】央行行长潘功胜出席G20财长和央行行长会议，强调继续实施好适度宽松的货币政策；CrowdStrike与OpenAI扩大合作；港股通买入Minimax规模超越阿里巴巴和腾讯”, by 钛媒体官方账号 — 钛媒体, 2026-09-02T23:20:00.000Z (ARR)
• Source URL: https://www.tmtpost.com/8126299.html