Best practices for AI agent tools
Short, practical guides on using MCP servers, agent skills and coding agents well and securely. Each one starts from a real incident or a real question, with steps you can copy and a checklist.
- Secrets and tokens
Move Trading API Keys for Vibe-Trading Out of Plain-Text Config
You will take the TRADING_API_KEY and TRADING_API_SECRET values out of the plain-text mcpServers config and load them from a secret store at launch. This lowers the chance that a trading credential leaks through a shared or committed config file.
AgentArmor team - Permissions and access
Confine a terminal-running MCP server to a limited folder and account
You will run DesktopCommanderMCP under a restricted account limited to a single project folder. This reduces the damage a misused terminal tool can do.
AgentArmor team - Secrets and tokens
Pass TrendRadar API keys as environment variables, not in Docker commands
You will run TrendRadar in Docker without typing API keys into the command line or committing them to files. Keeping secrets out of shell history and version control lowers the chance they leak.
AgentArmor team - Permissions and access
Give each swarm agent its own narrowly scoped API key
You will split a shared API key into one key per agent, each limited to the access that agent needs. This reduces risk if one agent is misconfigured or misused, since the sources flag improper scoping as a credential risk in multi-agent swarms.
AgentArmor team - Prompt injection
Review generated UI components and pasted text before running them
You will check AI-generated or pasted UI component code for hidden instructions and risky code before it runs in your project. This reduces risk from prompt injection, where text in a component or a prompt tries to make your assistant do something you did not ask for.
AgentArmor team - Permissions and access
Give an Argo CD MCP server read-only access by default
You will set up the MCP server so it reads Argo CD application state without the rights to trigger syncs or change deployments. This reduces risk if an agent misbehaves or is manipulated.
AgentArmor team - Installing carefully
Verify an npm package's publisher and source before running npx
You will check who publishes an npm package and where its source lives before running it with npx. This takes a few minutes and lowers the chance of running a look-alike or mislabeled package.
AgentArmor team - Permissions and access
Limit Apify tokens to the actors your MCP agent needs
You will review which actors the agent truly needs and keep the MCP server away from your broadest account token. This reduces risk if a token leaks or scraped content steers the agent.
AgentArmor team - Secrets and tokens
Check Credential Handling in mcp-use Before Connecting Real Accounts
You will confirm the package source and test credential handling with a throwaway token before connecting real services. This reduces risk of token exposure, since the directory entry flags credential exposure as a key risk.
AgentArmor team - Secrets and tokens
Move your Firecrawl API key out of shared MCP config files
You will find where the Firecrawl API key appears in shared MCP configuration and move it to private storage. This matters because a key sitting in a shared file can be copied, committed, or exposed, and a leaked key can drain your quota and billing.
AgentArmor team - Monitoring and response
Confirm the Pi MCP adapter does not log or cache your tokens
You will test whether the pi-mcp-adapter writes your credentials to logs or caches using a throwaway token before trusting it with real keys. This reduces risk of credentials leaking into files you might not review.
AgentArmor team - Configuration
Check MCP server versions and run one test before relying on it
You will confirm which versions of the GitHub and Playwright MCP servers your client loads, update them, and run one small test call before an agent depends on them. This catches stale or regressed tools early.
AgentArmor team - Supply chain
Verify your installed MCP package matches its published source before relying on it
You will confirm which MCP package version actually runs and that its installed code matches the source it claims to come from. This matters because published packages and repository source have drifted apart in real cases.
AgentArmor team - Supply chain
Verify which extension version actually fixed your checkpoint problem
You will check the Roo-Code log for the nested-repository message, remove or move the offending repo, and confirm checkpoints return. A version change alone does not prove the problem is fixed.
AgentArmor team - Supply chain
Pin Claude Code to a Reviewed Version and Review Each Bump
You will pin your Claude Code install to a version you have checked, so updates happen on purpose rather than silently. This reduces the risk of surprise changes that threads have reported in recent releases.
AgentArmor team - Supply chain
Pin Claude Code to a known-good version and stop silent updates
You will stop the Gemini Code Assist extension from updating itself onto a build that fails to start on older glibc systems. You will also protect Claude Code config directories before you change anything.
AgentArmor team - Supply chain
Confirm the pinned SDK version before upgrading your MCP adapter
You will check which MCP SDK version your adapter resolves before upgrading, so you can tell whether an OAuth issuer check will break your login. This avoids surprises from silent dependency changes.
AgentArmor team - Configuration
Run Continue MCP servers from WSL with a bash wrapper config
You will add a Continue MCP server entry that launches through WSL bash instead of a Windows executable, then test each server on its own. This matters because the default launch path failed for WSL workspaces in the thread.
AgentArmor team - Configuration
Keep one MCP config source and confirm which file each tool reads
You will pick one place for MCP server definitions and confirm which file Cline and Continue actually read. This keeps the two tools from drifting apart with different server lists.
AgentArmor team - Supply chain
Pin one zod copy that matches your MCP SDK to avoid type errors
You will align your project so only one zod copy is installed and it matches the version the MCP TypeScript SDK uses. This reduces the risk of registerTool schema type errors caused by duplicate zod installs.
AgentArmor team - Supply chain
Confirm your MCP SDK version and wire format before debugging transport errors
You will check the installed mcp version and whether your server exposes the session headers the Inspector needs. Ruling out version and CORS problems first keeps you from chasing transport bugs that are not in your code.
AgentArmor team - Supply chain
Check your installed MCP SDK version before trusting a documented fix
You will confirm which MCP SDK version your project runs and pick the fix that matches it. This keeps you from chasing workarounds for bugs that are already patched.
AgentArmor team - Monitoring and response
Diagnose a stalled MCP session before you resume the agent run
You will find out whether a stalled MCP session comes from an SDK version mismatch, a session routing problem, or a stale auth header before you restart the agent. Ruling these out first keeps you from resuming a run against a broken connection.
AgentArmor team - Configuration
Plan Around Rate Limits When MCP Tools Fire Parallel Requests
You will cap how many MCP tool calls run at once and keep their outputs small, so one burst does not stall or overflow the agent. The steps are general planning advice, since the threads I checked do not document a rate-limit error.
AgentArmor team