My AI Speaks `curl` and `jq` — MCP Can Catch Up Later

I asked my AI buddy what would help it dig in my DB. It arrived with a daemon, two port collisions, and a success message that only works on Gary’s Intel Mac — also known as “MCP server installation instructions.” So I taught it a simpler habit: Markdown guides + tiny curl/jq scripts. I get answers; my AI learns its tools; MCP can join when installs stop feeling like escape rooms.

A quick vibe check

You could build your own MCP server, park it in a private repo, and feel accomplished — until you need everyone else to use it. Now you’re writing install notes for mismatched runtimes, opening ports — wait a minute. Congrats: you’re Head of Ports & Feelings. New job description includes explaining why your MCP only works on :4312.
Are MCPs impressive? Yes. Lightweight for teammates and AI? Not really.

Here’s the deliberately boring alternative

A short guide in the repo.
And a handful of tiny scripts that do nothing fancy except work.

No daemons. No discovery layer. Copy, paste, run. AI follows the same guide; humans can audit every line.
All the guides are generated by AI, for AI.

The stack, in layers

• Data Provider’s API
• curl + jq boilerplate scripts — hide headers/auth/quoting; emit raw JSON
• Guides — AI-runnable Markdown with tenant/time-scoped recipes
• AI — reads the guide → runs scripts → pipes JSON through small jq formatters
• Human — skims the summary, clicks links, enjoys blissful ignorance

Humans struggle with fussy detail; AI thrives on it. Give AI small, deterministic commands and it assembles complex workflows while you sip coffee.

A repo shape AI can walk without tripping

repo/
├─ guides/
│  ├─ dev.md                   # developer's guide
│  └─ testing.md               # conventions, guidelines, examples
├─ integrations/
│  ├─ metabase.md              # read-only DB access
│  ├─ kafka.md                 # events investigations
│  ├─ linear.md                # Linear access
│  └─ coralogix.md             # logs, traces, and metrics
├─ scripts/
│  ├─ run_metabase_query.sh    # queries runner
│  ├─ logs_query.sh            # log search wrapper
│  └─ tickets_graphql.sh       # ticket system GraphQL wrapper
├─ formatters/
│  ├─ bi_summarize.jq
│  ├─ logs_top_errors.jq
│  └─ tickets_summary.jq
├─ playbooks/
│  ├─ investigate_ticket.md    # relevant data collection, code look-up
│  └─ incident_triage.md
└─ README.md                   # safety, quick start, conventions

How to help your AI help you

Prompt

Create a repo with `guides/`, `scripts/`, and `formatters/`. Each script wraps a service with `curl`, outputs raw JSON. Each guide is runnable Markdown. Each `jq` file turns JSON into skinbags-readable format.
Search online (or ask your AI) how to access each tool’s API and which environment variables to set.

A real AI run (obfuscated, because we’re civilized)

Prompt

Prompt
Using `integrations/metabase.md`: collect organizations (anonymized) and how many banks each has.
Then, using `integrations/coralogix.md`: compute how many logs each org generated in the last month.

Result

organizations=redacted, avg_active_banks≈redacted, logs_30d≈redacted

Timeline

• Discover guides & scripts — ~2 min
• Skim conventions & prep env — ~6 min
• Run queries & log counts (parallelized) — ~9 min
• Join, anonymize, export, summarize — ~5 min

Total: ~22 minutes. AI did the meticulous parts; the human was on a toilet break.

Closing

MCP might solidify one day. When installation and discovery get consistent across vendors, I’ll happily wrap these same HTTP calls behind it. Until then, a Markdown page and a handful of tiny, strict scripts are the shortest distance between “what the AI should do” and “the AI did it.”
Small pieces. Loosely joined. Shipped today.