Туториал по Codex: от базовых функций до goals, subagents, automations и скрытых механик

Summary of the video (Codex tutorial: basics → goals/subagents/automations + “hidden” mechanics)

1) Model/ecosystem context + product lineup

• The speaker claims Codex workflows used to rely mainly on “CLD…” model series, but OpenAI GPT-series models improved and now outperform in many coding tasks.
• They switch to OpenAI for ~90% of coding tasks due to:
  • better result quality
  • (they claim) fewer token-limit issues
• Mentions an “ultra-convenient” Kox App (Feb 2026) that combines:
  • CLI-like power (command-line tooling)
  • Desktop GUI convenience
• The tutorial covers the Codex ecosystem across multiple surfaces:
  • Codex App
  • Codex CLI
  • Codex (Exc?) / another execution mode (treated as part of the same system)

2) Core concept: local agent execution (not just cloud chat)

A key differentiator from “basic GPT chat” is that the agent runs locally on the user’s computer.

• The agent can:
  • create files in the local project folder
  • execute commands on the machine
  • set up/check results via a local server
  • use a built-in browser control to test functionality visually
• Prerequisite: you must install the Kox/Codex App locally; browser-only usage won’t work.

3) Pricing/review-style guidance + third-party sponsor mention

• Subscription guidance (as described):
  • start with any plan
  • for “normal development,” recommend a higher tier (~$100)
  • if limits still aren’t enough, a larger tier (~$200; claims 4× limits for 2× cost)
• Sponsor: Lessy AI (AI people search engine)
  • Workflow:
    • provide a text description of who you want
    • service asks clarifying questions
    • returns ranked profiles with links/contacts
  • Offers:
    • CRM export
    • personalized outreach email generation

4) Access control nuance (“permission gating”) + Auto-Review mode

• Default behavior: when new console commands/tools/parameters appear, the system often pauses execution until the user rejects/approves.
• Safety dilemma described:
  • stall on repeated command failures, or
  • grant broad “unlimited” permissions (less safe)
• Proposed solution: Auto-review access mode
  • If a command needs human review, a dedicated review agent checks it.
  • If rejected, the agent:
    • finds a safer workaround, or
    • escalates to a human review.
• The speaker reports using auto-review ~90% of the time, especially for normal work, but sometimes disabling it for “dangerous” mods.

5) Built-in vision/image generation for game asset creation (practical demo)

Using a Counter-Strike 1.6 modding example:

• The agent iteratively improves visuals:
  • adds textures (sand textures appear after enabling/setup)
  • fixes issues like crooked models, missing legs, camera/viewport alignment
• Image generation mechanism:
  • uses a built-in “Codex ability” to generate images (claim: it uses a GPT Image 2–like model)
  • supports reference images via vision abilities
  • mentions typical upload formats (e.g., “DragonRope files and screenshots”)
• Limitation + workaround:
  • the latest image model can’t generate transparent background images
  • suggested workarounds:
    • chroma key
    • third-party background removal
  • example: agent uses chroma key successfully
• Notifications:
  • enable completion notifications in Codex settings to avoid continuously monitoring (widget/app/Mac notification notes)
• Speed/quality tradeoff:
  • main runs use “Model 5 at High Era thinking level” (most powerful but slowest)
  • faster alternatives:
    • lower thinking levels
    • other fast OpenAI models (example mentioned: GPT “53X Spark”)
  • AST mode increases speed but consumes ~100% more tokens (i.e., doubles token usage, per speaker)

6) Vision/testing limitations + tools for visual QA

• Claimed limitation of current LLMs:
  • weak at accurately verifying “what looks good,” catching UI glitches, or judging visual correctness reliably
• For visual testing:
  • uses a browser use tool to load the running local server in a controlled browser and inspect UI/screenshots
  • example: checking a button in a test bench
• Browser automation plugin:
  • mentions a Chrome plugin enabling Codex to control the browser when authorized
  • example: checking Google Trends, but parsing/requests can trigger heavy blocks

7) Automation + scraping reliability: proxies (sponsored)

• For stable web automation, they recommend proxies.
• Claim: common proxy providers have many “dirty” IPs, causing redirects/captchas.
• Sponsor: Notmav(e)n / Nandmanzhe proxy provider
  • claims ~95% clean IPs
  • session lasts up to ~24 hours
  • location filtering + quality guarantee
  • crypto/card payments + MIR card support (Russia mention)
  • cashback + promo codes (discounts based on proxy type)

8) “Computer Use” for native app UI automation (and its limits)

• Codex can use a Computer Use tool to manipulate the computer UI outside normal Codex app flows.
• Demo: attempting to run “Everlasting Summer” on macOS
  • works only reliably for specific store/app versions (Steam version control issues)
  • agent gets stuck on complex UI flows (location map, card game)
  • context eventually overflows and control becomes unreliable
• Conclusion:
  • good for simple button automation
  • unreliable for long visual flows

9) Building a “personal Jarvis” overlay assistant (architecture + coding walkthrough)

The assistant:

• appears as a small overlay widget on top of macOS windows
• opens a conversation when clicked
• uses Gemini Flash 3.1 for voice/speed (speaker preference)
• can optionally fetch CRM data
• has access to Google Search tool
• runs Codex code in an interactive mode in a new tab/terminal for continued manual prompting
• continues chatting asynchronously while code runs
• disconnects when the user ends conversation

10) Project organization + AgentsMD / rules refactor

• They clean up the project folder and reorganize:
  • Services/ contains projects
  • System/ contains scripts/utilities
• They generate an AgentsMD file:
  • Codex uses it as the main “rulebook” for each session
• Refactor:
  • move detailed instructions into separate Markdown rule files under Rules/
  • AgentsMD points to where the rules live
• Convenience:
  • use a “dog symbol” (@) to reference files/folders with auto-complete

11) Git as a safety backbone (critical workflow)

• They emphasize initializing and using Git heavily:
  • keep multiple independent repos (one per service project)
  • root git ignores service folders to avoid mixing code
• Commit checkpoints before risky automation changes
• Branching/versioning approach:
  • revert to prior states if something breaks
  • states can be branched into “several independent versions” for parallel development

12) MCP servers: extending the agent with tools + documentation ingestion

• Integrate MCP Context 7:
  • adds commands to search for libraries by keyword
  • fetches detailed documentation for libraries
  • useful for macOS/native Apple features
• How MCP differs from:
  • Skills (instructions + resources)
  • Plugins (bundles of MCP tools + skills; e.g., Gmail plugin)
• Example: Gmail MCP plugin
  • reconstructs travel history by searching email for ticket/hotel data
  • used for visa-like requirements; claims it worked extremely well

13) CLI Codex integration + access mode behavior

• Uses CLI Codex (“console version of Codex”):
  • same capabilities, easier integration with other systems
• Access rights notes:
  • automation may default to “execution rejected” unless configured
  • ability to enforce auto-review via command prefix
• Sync between CLI sessions and Codex app:
  • requests can appear in the app chat history

14) CLI/App Server bridging to cover full Codex features

• They improve the widget/assistant by shifting from purely CLI execution to Codex application server calls.
• Claim: this enables 100% coverage of Codex features, showing desktop app sessions alongside widget control.

15) Goal/Gold + task management + subagents (more “agentic” reliability)

• Adds “goals” concept using an experimental feature named “Gold”:
  • enable via a Codex terminal command (example mentioned in subtitles: “Codex features enable a calls”)
  • pinned goal makes the assistant continue trying until it believes it’s done
• Limitation:
  • without explicit result-checking, completion bias remains (agent may stop early)
• Solution: subagents for a hierarchy of implementation + testing:
  • development agent implements
  • testing agents verify
  • completion happens only when tests pass
• Notes:
  • subagents are not used by default
  • mentions configurable “allowed depth” for spawned subagents

16) Automations: scheduling recurring tasks + error handling

• Demonstrates the Automations section:
  • view running automations, prompt, project, model, frequency
• Two creation modes:
  • in chat (ask the agent to schedule recurring work)
  • manually (choose prompt/project/model/frequency)
• Example automation:
  • every 10 minutes, check for a large video upload on a website
  • if something fails, agent handles it “calmly” and logs errors
  • contrasted with brittle “soulless” scripts

17) Extra “hidden mechanics” hinted

• Planning mode
  • agent is forced to plan first, ask clarifying questions, and only implement after plan approval
• Hooks
  • link scripts to events (e.g., run tests after development completes)
• Sleep prevention
  • “Prevent Sleep While Running” setting
• Mobile control
  • remote control of local Codex via phone app; session/chat visibility

---

Main speakers / sources

• Main speaker: Oleg (host; “my name is Oleg”)
• Primary product/tool sources referenced:
  • OpenAI models (GPT series, image generation)
  • Codex / Kox App / Codex CLI / Codex server (Codex ecosystem components)
  • MCP Context 7, Gmail MCP plugin, computer/browser control tools (as described by the host)
  • Lessy AI (sponsor)
  • Notmav(e)n proxy provider (sponsor)