Brex’s Jim automates three recruiting workflows with OpenClaw

Pedro Franceschi describes Brex’s virtual-employee approach using OpenClaw, including Jim, an AI recruiting employee, CrabTrap for HTTP-level agent controls, Magpby for token-cost analytics, and Pedro Franceschi’s markdown-based personal “autopilot” system.

• Jim has run since February and performs three recruiting functions: sourcing candidates, filtering inbound applicants, and providing recruiting analytics; it syncs Greenhouse, parses resumes, enriches LinkedIn profiles and GitHub data, evaluates candidates against role-specific criteria, and can source candidates weekly for a job posting.
• CrabTrap is open source and places an HTTP proxy around an agent to inspect network traffic; it applies static rules first, sends non-static requests through an LLM policy judge, and can generate a policy by replaying observed HTTP traffic—Pedro Franceschi said one blocked Granola request used Sonnet, took two seconds, and consumed about 1,000 input tokens and 104 output tokens.
• Magpby tracks AI usage across corporate AI, operational AI, and product AI; in the demo, transaction tagging cost about $0.10 per call and roughly $28,000 per month, while disputes cost about $2 per transaction, described as a 240x difference. Pedro Franceschi said Brex gives engineers an unlimited token budget with a few caps, then optimizes costs after adoption is established.

Peng Zheng built a self-updating personal check-in website with Grok Bot, Google Places API, image generation, and structured data, while John Bai used his Figma Bro bot through Figma MCP and voice memos to automate Figma production work and create interactive prototypes.

• Peng Zheng’s check-in bot accepts a photo, screenshot, or place name; it identifies the venue through Google Places API, retrieves location coordinates, processes the image, removes people, isolates the building facade, and generates consistent light-mode and dark-mode 3D-miniature images for his website.
• For a Bernal Heights Peak check-in, Peng Zheng submitted the location plus Claire Vo’s and John Bai’s social handles; the bot created the visual entry, captured the park swing, and published the result to the website.
• John Bai used Figma Bro with a Figma MCP connection to place and resize icons across bot assets from Figma screenshots while at the gym at 8:52, and later used a linked Figma template to generate bot-marketplace marketing materials; his DevBot/experiments workflow turned a voice-described mobile sharing interaction into two working prototype options.

A Grok Bot team of 11 specialized bots is configured to coordinate through shared channels and scheduled tasks for email, planning, YouTube production, website analytics, X research, money-saving, family logistics, fitness, digital cleanup, and school-related errands.

• Dr. Light was imported from Lauren Tan’s marketplace bot, “Dr. Eggbot,” then renamed; it creates bots such as a joke bot named Punchline and runs scheduled checks to improve other bots’ prompts, skills, and access.
• Chief reads email and calendar items, drafts replies, accepts invitations, and delegates work to other bots through its remote computer and browser; it assigned Frugal Dad to list headphones on Facebook Marketplace and Loving Husband to check a children’s school newsletter on weekdays at 4 p.m.
• YouTube Producer runs every Monday and Wednesday at 7 a.m. to collect 7-day channel outliers from Peter’s and similar YouTube channels, generate video ideas, and create thumbnail/title packages, hooks, outlines, and Linear tickets; Health Coach separately reports every Saturday at 8 a.m. using Withings smart-scale weight/body-fat data and a workout app, including one week with 27,000 pounds lifted across five sessions.