Building AI Apps Live from Cairo 🔴
Want to back the stream? The «داعم» membership on YouTube is EGP 60 a month: badges, emoji, and I answer your questions first on stream. Become a supporter
The gist
The stream documents practical work on a website interface and a project command center, alongside broader discussion of coding-agent workflows. UI reviews covered seamless animation, cards, locked states, legibility, and Telegram calls to action, emphasizing that good visual results require explicit standards, reference screenshots, and human review. Some visual effects still needed refinement, while state automation was deferred because the future flow and requirements were unclear. For the command center, the discussion covered dependencies, waitlists, Shopify requirements, and human, agent, and paired execution modes. A custom MCP server was then built to provide task context and receive status updates: basic create, edit, and delete tests worked, while Electron refresh behavior and fuller milestone and Agent Hub coverage remained unresolved. The stream also explored enriching agent instructions with current context, acceptance criteria, and references; reviewed an agency website from visual and sales-funnel perspectives; and considered turning live-build lessons into courses if audience interest is demonstrated. Livestream stability, YouTube API behavior, and Cloudflare deployment were monitored throughout; some states later appeared correctly, but several underlying causes were not conclusively established.
53 moments from the stream
- 2:13:43
Closing MCP Gaps in Milestone and Task Management
A review of command-center data that the MCP tools could not yet fully access or update, including milestone exit criteria, actual dates, schedule drift, task properties, and task history. The discussion proposes new read/write tools, automatically stamping a milestone when its first task begins, and deriving drift from planned versus actual dates instead of maintaining it manually.
- 2:21:53
Agent Hub Coverage and MCP Client Compatibility
An examination of what the MCP tools cover in the agent hub and what remains missing, such as reading the full activity feed, listing connected agents, performance statistics, and registering a new agent. According to the response shown during the session at that time, MCP-compatible clients could connect directly, while clients or models without MCP support would need another transport such as a REST API built around the existing read/write logic.
- 2:28:07
Testing Command Center Control Through MCP
The presenter verified that the new MCP server could read Command Center state, create a test card, change its priority and execution mode, and then delete it. The interface also had a refresh limitation: changes appeared only after reopening the application, and the presenter proposed adding a refresh mechanism when the Electron app is deployed.
- 2:34:34
Exposing Command Center Tools to Codex Through the CLI
The discussion explores giving Codex capabilities similar to those Claude receives through MCP, but via command-line tools documented in the prompt or project instructions. The example describes Codex retrieving context and updating status with results printed as terminal text; the tool or project name is not fully clear in the captions.
- 2:36:22
Building an Arabic AI-Coding Learning Community
The presenter invited viewers to help grow a small Egyptian Arabic Telegram community for people interested in AI-assisted coding and business, with an emphasis on learning from one another and developing shared expertise.
- 2:37:29
Troubleshooting Livestream Stability
The presenter explained that an earlier stream disconnected despite a stable wired fiber connection and attributed the problem to the ISP. He lowered the streaming bitrate, adjusted some settings, and then opened the stream analytics.
- 2:41:04
Testing Agent-Driven Task-Board Edits
An agent was asked to add a test card with P3 priority and agent execution mode, after which the task board was reopened to verify where the card appeared. When the task-search result did not match the intended task, deletion was requested, illustrating the need to verify agent-made changes in task-management tools.
- 2:42:35
Codex MCP Support and Connecting Two Agents to One Server
A question was raised about whether Codex uses MCP and what its alternative would be if it did not. After checking, the discussion concluded that Codex supports MCP, and two agents were configured to use the same MCP server so their access to command-center data could be compared with the same query.
- 2:50:21
Reviewing an Agency Website and Brand Identity
A largely positive walkthrough of a digital agency site covering its clear navigation and hero, work showcase, interactions, testimonials, calls to action, FAQ, and footer. The speaker qualified the review as a subjective opinion rather than a web developer’s assessment, and suggested reconsidering the common purple AI aesthetic only if it is not an intentional part of the brand.
- 2:55:22
Turning a Service Website into a Marketing and Sales Funnel
The discussion shifts from visual quality to client acquisition: strengthen proof, fix broken calls to action, reduce CTA confusion, and offer a lower-commitment way for leads to leave details before booking a meeting. The suggested sales process depends on price: higher-priced services may justify discovery calls, while lower-priced offers need a faster, more automated path. Once the site is ready, the proposed priorities are visibility, direct selling, refining the service, and examining profitability and the marketing funnel, potentially with an LLM as a planning aid.
- 3:04:56
Improving Coding Models with Context, References, and MCP Access
A discussion of connecting Claude and Codex to tools through MCP, with the observation that models tend to perform better when given deep context, precise instructions, and reference material such as SVG files or screenshots. It then introduces reviewing prompt quality before asking models to execute tasks in a work system; these pages do not show or assess that execution.
- 3:10:15
Turning Weak Task Labels into Context-Rich Agent Briefs
The discussion explains why a short task label is insufficient: an agent may have to guess the template, configuration, expected output, and success criteria. It proposes enriching each task at execution time from the current codebase and project history instead of processing every task once in advance. Explorer and planner agents could assemble the prompt, acceptance criteria, constraints, context files, and references, potentially through an MCP task-preparation tool using the existing step-runner and child-process pattern.
- 3:21:59
Checking a Livestream After a Brief Interruption and Resume
After an issue in the streaming software, the YouTube connection was resumed shortly afterward. The stream status was then checked, with no dropped frames reported, before the session was closed.
Got a question about something in this stream?
Ask in the group