Gecko — the API comprehension layer for agents. Point an agent at any API — even one behind human-shaped docs and a paywall — and it finds the right call, makes it correctly the first time, and runs. No client to write, no guessing whether the agent is calling it right.

Docs and endpoints are built for humans. Gecko translates an API's surface into question-shaped, first-call-correct agent tools — and handles the auth handshake the docs don't explain.

The engine — gecko-surf (open-source, on PyPI) — turns any OpenAPI into first-call-correct MCP tools:

curl -fsSL https://get.geckovision.tech | bash
gecko <openapi-url>          # → comprehension summary + a one-click "add to Claude/Cursor"

I'm running Dev3Pack AI Engineering right now — a three-week cohort, fifteen live sessions, a self-paced Python fast lane and a capstone, 14 Sep to 2 Oct 2026.

It is graded the same way I argue about the day job: every scored exercise runs offline against a deterministic fake model, so no key, no network and no bill stands between anybody and finishing. A submission is a claim plus the notebook that produced it — the tool refuses to hand in a notebook nothing was run in.

Building a course in public is the cheapest honesty test I know: every claim gets run by thirty people on thirty different machines, the same week you make it.

Enrolment is closed for this cohort. Ask me about the next one.

A coding agent can one-shot a popular API. It can't one-shot the Nth painful one — long-tail, messy, poorly-documented, often paywalled. It picks the wrong endpoint, misplaces the credential, sends a UI amount where atomic units are required, forgets the idempotency key. Each miss is a failed call, a retry, a burned credit. Gecko replaces guess-and-retry with first-call-correct comprehension — and stays correct as the API drifts.

gecko-surf — the comprehension engine. Ingest an OpenAPI surface → question-shaped, first-call-correct agent tools (auth hidden, units / placement / idempotency resolved), served over MCP. A $0 recorded mode simulates the call offline before a cent moves.

ayuda-venezuela-bot — a Spanish-first Telegram bot giving Venezuelans clear emergency information, an agent over the SOS Venezuela API. Comprehension pointed at something that isn't a demo.

The three verbs. APIs get PAID (rails), skills get DISTRIBUTED (marketplaces), APIs get USED (Gecko — comprehension). We compose on x402 / MCP / pay.sh — not a payment rail, not a marketplace.

stack = {
    "engine":    ["Python 3.11+", "uv", "MCP (Streamable-HTTP)", "stdlib-first", "ruff", "mypy"],
    "chain":     ["Solana", "Anchor", "Pinocchio", "IDL comprehension", "surfpool forks"],
    "delivery":  ["PyPI: gecko-surf", "one-click MCP for Claude / Cursor / VS Code", "Vercel", "AWS ECS"],
    "teaching":  ["Jupyter", "deterministic offline grading", "GitHub Actions as the gradebook"],
    "model":     ["open-core — open engine, private correctness corpus"],
    "compose":   ["x402", "pay.sh", "OpenAPI 3.x"],
}

How I work: every wire integration ships a free offline simulation first, so it can be falsified without spending anything. Live smoke is the last check, never the debugger. A number that wasn't measured doesn't go in the README.

• Dev3Pack cohort through 2 Oct — fifteen sessions, a week at a time
• First 10 outside-network teams running gecko-surf on their Nth painful API weekly
• The correctness corpus — every call teaches how to call API X right
• Surfpool / Solana-AI-Kit — comprehend the chain's tools and the APIs around them

Patches I sent to projects other people built and maintain:

• Orquestra, by @berkayoztunc — his project, IDLs into APIs and MCP servers. I sent two fixes: one so the signer refuses bytes that are not the ones that were checked (verification is worth nothing if what gets signed can differ from what was verified), and one so a transaction reports its own compute units rather than an inner CPI's.
• DeerFlow, by ByteDance (76k★) — I contributed the Portuguese (pt-BR) README, so Brazilian devs can pick it up in their own language.

Mine, and open:

• gecko-ai-coach — a grounded course coach that ships with its own hit rate, and a contribution bar: a pull request has to carry its before-and-after number.

Building agents that have to call real, painful APIs? That's the whole job.

DM me on X or LinkedIn