Boundary Map
Every tool, transport, credential, external API, and write action mapped into read-only, write, destructive, and privileged paths.
48-hour async review for teams shipping agent tools
I review one agent, MCP server, or tool-using product for the practical failure modes that break launches: tool boundaries, secrets, auth, prompt/tool injection, test gaps, and deployment assumptions.
What ships
Every tool, transport, credential, external API, and write action mapped into read-only, write, destructive, and privileged paths.
Ranked issues with reproduction steps, evidence, affected files, severity, impact, and the lowest-risk fix path.
Concrete tests for tool contracts, auth gates, secret redaction, retry behavior, and failure states that agents commonly trigger.
Short handoff covering what to fix now, what to monitor, and what to defer without increasing customer-facing risk.
Sample evidence
The sample audits cover three real public repos: douban-mcp, a public MCP server + CLI with auth, write tools, and external scraping; firecrawl-mcp-server, a production MCP server with hosted/local transports, OAuth, monitor tools, and local parsing; and browserbase/mcp-server-browserbase, a browser automation MCP server with stdio and HTTP transports, sessions, page actions, observation, and extraction tools.
Open the douban-mcp sample report Open the Firecrawl MCP sample report Open the Browserbase MCP sample report Compare all three sample reportsBest fit
Free triage tool
The repo includes a browser scanner for public GitHub URLs, a private local-file scanner, a Node script, and a reusable GitHub Action that scan an agent or MCP codebase for review signals: transports, write actions, credential paths, auth gates, redaction, tests, and CI.
Open the MCP server security scan page Open the AI agent security audit page Open the AI Agent Security Radar Open the MCP Security Radar Use the GitHub Code Scanning workflow Paste a public GitHub URL Use the scanner Open the MCP security checklistReady to start
Use this local-only builder to prepare the exact scope, delivery preference, and payment network before opening a GitHub request. Public GitHub repo requests get an automated no-execution scanner triage comment before paid scope acceptance.
Payment
Open an audit request, include the repo/product slice, and pay after scope is accepted. ETH, ERC-20 USDC/USDT/DAI, SOL, and SPL USDC are ready now; if you need an invoice-first discussion, choose that option in the intake form.
Review the Statement of Work before payment Open the fixed $1,000 quote0xa7F2235a77FBc4eCcbF60923BCDF6Df74eC710FF
Accepted assets after scope acceptance: ETH or ERC-20 USDC/USDT/DAI.
5CjUaMAsbXx2Hjczwoqi4MChTU1KjfUzbdiwPqZeceVM
Accepted assets after scope acceptance: SOL or SPL USDC.
Operator
Backend engineer focused on Agent Harness, Tool Use, MCP, Evals, and AI infrastructure.