Coding by Voice: Goodbye Keyboard with Codex
Craig Hewitt demonstrated how he controls his computer entirely by voice using Codex's voice mode.
1What is Codex's Voice Mode?
Codex's voice mode allows users to control their computers entirely through voice commands, unlike traditional dictation tools. Craig Hewitt showed how he could manage everything from browsers to IDEs using this feature. The voice interface remains active even when switching between applications, eliminating the need to constantly return to the chat screen.
2Setup Process
To use voice mode, you first need to enable computer access in the Codex app. Craig recommended configuring browser access to require approval before the agent accesses most applications or sites. This is a critical step to prevent security vulnerabilities and unintended actions. Only trusted development sites are granted broader access permissions.
3Browser Control
Craig demonstrated how he could manage basic browser tasks like opening websites, searching for information, and navigating pages using voice commands. The agent can also automate routine tasks such as filling out forms when it already has the necessary context. This feature offers significant time savings, especially for repetitive tasks.
4Cross-Application Workflow
Codex isn't limited to browsers; during the demo, Craig showed how the agent could open Cursor, find a specific repository, and report uncommitted changes. After receiving permission, it could commit those changes. This proves developers can manage their entire workflow by voice.
5Adding Page Content to Chat
Craig demonstrated the ability to select a specific part of a webpage and add it directly to the chat. This provides the agent with the necessary context to act on a particular element, such as focusing on a specific interface section you want to modify.
6Security Measures
Browser and computer control can pose risks like unintended actions and prompt injection from web content. Craig emphasized requiring approval for most applications, granting broader access only to selected tools and local development sites, and avoiding untrusted sites. This approach is crucial for maintaining security.
7The Delegation Problem
The demo highlighted the delegation problem that arises when an agent can act on your behalf. Craig uses permission settings, a trusted sites list, and human review to manage these risks. Users need to decide which tasks they're comfortable delegating, especially for sensitive operations.
8The Future of Voice Control
Craig's demo showcased the potential of voice control to transform not just coding but entire computer usage. The next episode will feature Jayeeta Putatunda from Turing building an agent for financial analysts. This agent will categorize financial news and rank stories based on analysts' coverage profiles.
Original article: O'Reilly Radar — Sesle Kod Yazma: Codex ile Klavyeye Elveda