I direct AI agents to do real engineering work: research a problem, build the fix, verify it holds, and ship it.

247 pull requests merged: 223 in my own repositories, where the tools below live, and 24 in other people's projects across 16 projects, 5 of them bug fixes. I have sent 115 to other projects so far, 56 still open.

_{Counted from public GitHub data on 2026-10-24.}

Bug fixes merged by other maintainers, newest first:

• 2026-10-22 x402-foundation/x402#3826 fix(core): do not mutate the extensions object while building a 402
• 2026-09-10 cloudfoundry/terraform-provider-cloudfoundry#592 fix(provider): guard nil MaintenanceInfo on managed service instance create
• 2026-07-03 nexu-io/open-design#5078 fix(amr): normalize Vela JSON catalog ids on the live fetchModels path (fixes #4410)
• 2026-06-24 activepieces/activepieces#13806 fix(cal-com): migrate webhook registration from decommissioned API v1…
• 2026-06-19 windmill-labs/windmill#9680 fix(home): preserve URL filter state on page refresh

Open now in other people's projects, code and spec changes, newest first:

• 2026-10-23 Chain-Love/chain-love#4786 Add Bug Bounty Intelligence (provider, MCP server and security scanner on Base)
• 2026-10-23 tigerless-labs/autoharness#278 fix(redact): redact bare vendor-prefixed keys and secret-named variables
• 2026-10-16 x402-foundation/x402#3830 spec(offer-receipt): add the svm-memo authorization method
• 2026-09-28 x402-foundation/x402#3604 specs: propose Buyer Delivery Claim extension
• 2026-09-15 inercia/MCPShell#54 Fix shell-quote injection in the weather example, add a shellQuote template helper
• 2026-09-15 cnosuke/mcp-command-exec#1 Validate arguments, not just the program name, in IsCommandAllowed

_{The numbers, the two lists above and the project list below refresh on their own from public GitHub data, see data/evidence.json.}

• Agent-commerce trust infrastructure: tokenizen, signed delivery receipts for AI agents over x402, live on Base mainnet. Other engineers have checked it from the outside: a clean-room Python verifier by goun7 matched all 7 test vectors and now runs on every push in Tamga, StelarDigital ran the metered vector 12 of 12 green, a protocol maintainer linked our verifier in his official docs, and two developers built on the claim format in asm-spec and again. A draft ERC-8004 convention takes its metered delivery block from our spec. All the evidence.
• x402 protocol itself: found and fixed a bug in the core library where building a 402 response overwrote part of a route's own extensions declaration. Merged by the maintainers.
• x402 receipt verification: agent-receipt-verify, one command checks whether a paid endpoint follows the x402 spec, no account, no key. Benchmarked against 60 real sellers from the Coinbase index, with the method in the repo.
• AI agents running real infrastructure: al-yad is a free open-source browser agent for your own Chrome, put through an 18-agent adversarial code review and attack runs against the real running code before release (write-up). muraqib opens its own fix PRs when nightly tests break, and its work surfaced a gap in Playwright's CI guidance that a Playwright maintainer fixed the next day.
• AI-agent security research: flagged 22 AI-agent repositories for a CI bug class that leaks secrets. 3 confirmed and fixed, among them a High-severity GitHub Security Advisory with credit, plus a High-severity report accepted by GitHub for its own infrastructure. Details below.
• Smart contract security: bug-bounty-intelligence-mcp and al-mizaan-judge, built on 1,032 findings from 10 Sherlock audit contests.
• AI-driven health platform: Longevity AI, live for the Dutch market, built solo, roughly 190,000 lines of code.

• bug-bounty-intelligence-mcp: MCP server that scans Solidity repos through a 7-gate verification framework (Al-Mizaan) to cut LLM false positives, benchmarked against Slither; free pattern-search... Latest release v1.0.3, 2026-07-29. Updated 2026-10-24.
• tokenizen: Lets AI agents record, in a way anyone can check, whether they actually got what they paid another agent for over x402. Cites and reads the agent-economy stack it... Latest release v0.5.0, 2026-09-06. Website. Updated 2026-10-24.
• muraqib-agent-watch: Free, instant capability check for any MCP server: what can it actually do, and does it say so Updated 2026-10-23.
• agent-receipt-verify: MCP server and library for independent verification of agent payment and outcome receipts. Written against the published specs, tested against the protocol authors'... Latest release v0.2.1, 2026-10-17. Updated 2026-10-19.
• al-yad: Yad: a free, open-source, local-first browser agent. Chrome extension (Part A, the hand) + local Companion (Part B, the brain). Your API key is encrypted on your... Latest release v0.1.2-companion-npm-standin, 2026-09-11. Website. Updated 2026-10-11.
• muraqib: Nightly Playwright QA that tells you when it stops watching. npx muraqib doctor finds the silent-cancel bug in any existing setup. Claude opens the fix PR. MIT, no... Latest release v0.3.4, 2026-09-29. Website. Updated 2026-09-29.
• claude-memory-trim: Keeps Claude Code session memory lean by rotating session logs between hot and cold storage. Cuts token cost at session start by 60-80%. Updated 2026-09-08.
• al-mizaan-judge: Strict 7-gate validity judge for a smart-contract bug bounty finding, a second opinion before you spend a Sherlock/Immunefi/Cantina submission. Updated 2026-09-08.
• 3ilm-mcp: MCP server: 1,032 verified smart contract vulnerability findings from 10 Sherlock audit contests Updated 2026-09-08.
• postmortems: Short technical postmortems: what looked right, what turned out wrong Updated 2026-08-24.
• automation-guardrails: Two small guardrails for Playwright/Puppeteer: verify you're on the right page, and fill contenteditable/rich-text editors without the silent double-write bug Updated 2026-08-03.

I check GitHub Actions workflows across popular open source AI-agent frameworks for the bug class that leaks secrets or lets a broken security check pass silently, and, more specifically, AI CLIs (Claude Code, Copilot CLI, Codex) wired directly into a job and triggered by public, untrusted input, with real secrets sitting in the same job. 22 findings across separate repositories this month, three confirmed and fixing, one escalated to a major AI team's own security group, six retracted by me after I checked my own assumption against the tool's actual source code and found it wrong.

Background: the original CI-hygiene piece and the full 22-repository breakdown.

• pymc-marketing: public GitHub Security Advisory, severity High, reported by email, fixed within a day, credited as reporter. Write-up.
• sheepworrier/BilliardsScorer: script-injection plus missing-author-check in a claude-agent.yml workflow, reported by email, both fixed exactly as reported within two days.
• github/request-marketplace-action: unquoted shell variable in a self-hosted-runner workflow allowed argument injection into a git clone call next to a live privileged token, reported through GitHub's private vulnerability disclosure process, severity High, credited as reporter (accepted), fix verified live in the current code. Advisory itself is not yet public, so no link here.

Open PRs: autogen, semantic-kernel, letta-code, adk-python, smolagents, promptflow, langflow, E2B, griptape.

Open issues: mem0, camel, goose, mastra, n8n.

Open to roles and freelance work in AI agent infrastructure and orchestration, plus CI/CD supply-chain security and AI-agent hardening. The security findings above were found on my own initiative, in passing. Email [email protected].

Longevity AI is an AI-driven health platform for the Dutch market: 28 questions in, a personal 6-month nutrition and lifestyle plan out, no diagnoses or medication advice. Built for people with chronic symptoms but no clear diagnosis, HR wellness programs, and holistic practitioners who want their own client portal. Live, roughly 190,000 lines of code, built solo, TypeScript end to end.

Runs on its own automation: muraqib watches it nightly and merges its own fixes, a content pipeline turns new health research into trilingual posts every few hours, and a weekly self-improvement loop scores synthetic patient runs to sharpen the next report. All of it costs a few cents a week.

🌐 Platform: longevityai.nl 📬 Contact: [email protected]