Imported from xnoto/codex-config (
AGENTS.md). Install upstream withnpx skills add xnoto/codex-config. Copyright stays with the author.
MCP routing
Select by the named target environment. If it is unspecified, ask before querying or changing anything.
Hatch resources: use only aws-staging, aws-prod, kubernetes-staging-eks, kubernetes-prod-eks, argocd-staging-eks, argocd-prod-eks, and grafana. kubernetes-staging-eks and kubernetes-prod-eks are fixed-context Kubernetes gateways for their respective Hatch EKS environments; they are distinct from makeitwork-kubernetes. These are separate servers, so their tools carry no extra prefix — Hatch Grafana is the grafana server's query_prometheus, while the Hatch Kubernetes gateways use the backend's own tool names such as pods_list, not kubernetes-staging-eks_pods_list.
Make IT Work Cloud resources: use the direct Make IT Work Cloud remote servers. Seven use bare integration names — apify, aws-docs, context7, parallel-search, playwright, slidespeak, and terraform-docs. Six use the makeitwork- prefix: makeitwork-argocd, makeitwork-aws, makeitwork-grafana, and makeitwork-kubernetes retain it because their bare names collide with client integrations that represent other environments; makeitwork-cloudflare and makeitwork-gcp use it to make the Make IT Work Cloud scope explicit. These names target only Make IT Work Cloud resources, not other environments; the prefix is a naming distinction only, not a security boundary. All thirteen are the same kind of remote server at https://mcp-<integration>.makeitwork.cloud/mcp, authenticating with CF-Access-Client-* headers referenced from the CF_ACCESS_CLIENT_ID and CF_ACCESS_CLIENT_SECRET environment variables. Each server exposes its backend's own tool names — makeitwork-grafana's query_prometheus, makeitwork-argocd's list_applications, makeitwork-kubernetes's pods_list. AWS itself is one such server, whose tools read makeitwork-aws's aws___<tool>; the server is not AWS-specific despite that prefix. github, hero-ssh, and codebase-memory have no external endpoint and stay on internal or client-local transports.
makeitwork-cloudflare maps to https://mcp-cloudflare.makeitwork.cloud/mcp; makeitwork-gcp maps to https://mcp-gcp.makeitwork.cloud/mcp. The makeitwork- prefix scopes the client-side name only; do not prepend it to the endpoint hostname.
Environment-neutral tooling also arrives through direct servers: parallel-search (web), context7 (library docs), aws-docs, terraform-docs, and apify. opentofu-docs is a standalone server. Each integration is its own [mcp_servers.<name>] entry again — bare for the seven, makeitwork--prefixed for the six — so enabled toggles it individually; the aggregate makeitwork entry no longer exists.
Tool usage
The load-bearing usage instructions are restated here.
Web (the parallel-search server): reach for web_search first for factual, current-information, research, comparison, and troubleshooting questions. Its excerpts are meant to be answered from directly — do not fetch every result. Pass multiple search_queries in one call rather than chaining. Use web_fetch only when the user names a specific URL, you need exact wording, or the excerpts conflict or are clearly insufficient.
Library docs (the context7 server): use for any library, framework, SDK, API, CLI tool, or cloud service question — API syntax, configuration, version migration, setup, library-specific debugging — even for well-known ones, and even when you think you know the answer, because training data lags. Prefer it over web search for library docs. Not for refactoring, business-logic debugging, code review, or general programming concepts.
AWS docs (the aws-docs server): search_documentation with specific technical terms, then read_sections when the table of contents localizes the answer, otherwise read_documentation. Paginate long pages with start_index. Fall back to recommend when repeated searches come up short. Always cite the documentation URL.
Grafana (the makeitwork-grafana and Hatch grafana servers): timestamps without a timezone offset are interpreted as UTC. Include an offset such as -05:00, or use relative syntax like now-1h.
Terraform/OpenTofu docs: query the registry for current provider and module versions before generating configuration, and pin what you generate.
Apify (the apify server): pay-per-event with real money, and it returns bulk datasets. Exhaust the web and docs tools first; reach for it only when the target is login-walled or anti-bot (Facebook Marketplace, Google Maps) or when structured listing records are the actual deliverable. Search the store first — a relevant Actor usually exists — prefer Actors with higher usage or ratings, always check an Actor's input schema before running it, and bound every run with result limits (resultsLimit/maxItems), price filters, and location radius.
Codebase Memory routing
codebase-memory is a local, derived code-discovery index served by the MCP gateway. Use it only for repositories below ~/git, and index each repository explicitly rather than indexing the parent directory. Keep shared graph-artifact persistence disabled so repository source is not modified. Indexing changes local derived state and requires confirmation. Its local index can be stale or incomplete; use GitHub for exact file reads, remote branch heads, repository writes, and freshness-critical claims.
