Background: Seeking a Stable and Cost-Effective Core for AI Agents
As software capable of autonomously understanding, planning, and executing complex tasks, the capabilities of an AI Agent are highly dependent on its underlying Large Language Model (LLM) API calls. For popular desktop agent applications like Codex, users typically rely on OpenAI’s GPT series models. However, using third-party relay APIs introduces stability issues like cache invalidation, rate limiting, and reduced context length, while direct use of official APIs carries risks of account suspension or high subscription costs.
Against this backdrop, a test was conducted to verify the feasibility of using a Chinese large model API. The test selected the ‘Step Plan’ service offered by StepFun, which uses the step-router-v1 model router to automatically switch between models like deepseek-v4-pro and step-3.5-flash based on task complexity, offering a solution that balances performance and cost. The core question of this test was: after integrating such a Chinese API into Codex, how much of its core workflow can be preserved?
Integration and Basic Functionality Verification
The test first resolved the API adaptation issue with an automation script. This script utilized the cc-switch tool to convert the standard StepFun API format to the Responses API format required by Codex and guided the user through the model configuration. Once configured successfully, the Codex application interface clearly indicated the model source as StepFun, signifying the switch to API mode.
On a basic functional level, the test verified two key interactive experiences:
- Mobile Device Connection: Even in API mode, users could still connect to and operate Codex from their mobile phones, including making calls from the lock screen.
- Core Shortcut Functionality: Codex’s core feature of capturing and analyzing the screen by double-pressing the Command key remained fully functional in API mode.
These two verifications demonstrated that API integration did not diminish Codex’s core capabilities as a cross-device, high-efficiency tool, laying the groundwork for more complex functional testing.
Core Workflow Testing: From Browser to Desktop Automation
The test delved deeper into Codex’s core automation capabilities, covering scenarios in web browsing, desktop interaction, and code generation.
Browser Automation: Using the @chrome command, the test instructed the model to post on Twitter in a browser with two different user profiles open simultaneously. The step-3.5-flash model initially got confused but was able to self-correct, identifying and switching to the correct browser user profile, and successfully completed the task after 3 to 4 rounds of exploration. This showcased good environmental awareness and error correction capabilities.
Desktop GUI Automation: With the @computer command, the test required the model to send a message to a specific contact within the Lark application. The model successfully drove the mouse and keyboard to perform a series of GUI operations, proving its ability to control applications across the desktop.
Code Generation and Tool Invocation: The test asked the model to write a Python script to call the image-edit-2 image model included in the ‘Step Plan’. The model not only successfully generated a usable script but also verified the feasibility of using API mode to call multimodal models for text-to-image generation and image editing.
Task Intervention: Users can intervene in real-time during task execution. The test showed that Codex in API mode could fluidly respond to both ‘guiding’ (correcting the current task) and ‘queuing’ (adding a new task after the current one), which is crucial for complex workflows requiring human-computer collaboration.
The Challenge of Long-Term Tasks and Memory
The practicality of an AI Agent lies not only in single-task execution but also in its ability to handle long-term, multi-stage tasks and utilize historical information.
Long-term Goal Setting (Goals): By enabling Codex’s goals feature, the tester set a long-term goal to ‘develop a gallery webpage for displaying AI-generated images’. The model was able to work continuously towards this goal and supported mid-process goal modification, showing potential for handling complex projects.
Persistent Memory: The test configured an Obsidian vault as the persistent context storage directory for Codex. The model successfully queried and retrieved data and file paths about Codex workflows stored in the knowledge base months prior, proving that its external memory reading capability remains effective in API mode.
Identified Limitations
Although the test results were largely positive, they also highlighted two functional limitations in API mode:
Remote Device Connection: Because it uses an API key for login instead of an account, Codex cannot directly connect to other devices under the same account that are on a non-local network (e.g., controlling a remote Mac mini from a MacBook Pro). Such operations still require traditional remote connection methods like SSH.
Chronicle Memory Component Disabled: Codex’s native Chronicle component, which automatically extracts and builds memory from the user’s recent screen activity, is disabled in API mode. This means the agent cannot passively learn the user’s immediate operational context in real-time.
In summary, the test of integrating StepFun’s API into Codex shows that Chinese large models are already highly practical for driving complex AI agents, covering most core scenarios from basic interaction to automation and long-term tasks. Although there are limitations in certain advanced features, the cost advantages and supply stability make it a viable alternative. It also provides an effective benchmark for evaluating the compatibility of other models and agent frameworks.
