agent-connect-prototypes

Demo/product experiments for the NEAR Intents Agent API ("Agent Connect") — the API that gives every AI agent its own agent account on NEAR Intents, governed by rules its owner signs once with their wallet: policies (actions, assets, limits, USD budgets, timelocks, schedules, approval, freeze), revocable per-agent grant tokens, and confidential balances.

The thesis: give an agent a pool of money it can transact under your direction or autonomously, knowing it will operate within guardrails you signed.

Browsing the experiments

The experiment content — the ideas themselves — is maintained privately and does not live in this public repo. The public layer is the machinery: the audit protocol, the installable skill, the site build script, and this process documentation. Propose new ideas by opening an issue or a PR adding a markdown file under docs/; the maintainers merge, move the content into their private workspace, and keep this repo clean.

The audit runs entirely in your local clone: open the repo in any AGENTS.md-aware harness, type an idea in chat, and the agent will run the gates, grill you in rounds, and write the verdict to docs/ locally.

The idea audit

Ideas arrive in chat and get audited hard through four gates before they touch docs/:

1. Relevance kill-test — would the idea work just as well with a plain API-key wallet? If removing the signed policies, grants, budgets, timelocks, approvals and confidential balances doesn't change the demo, it's not an Agent Connect idea.
2. Demo target — the 1-minute video beat-by-beat, the minimal demo surface, and the single moment the human sees the guardrail bite.
3. Aha test — one sentence a skeptical viewer would repeat. Generic "agents + crypto" fails; the rule biting live passes.
4. Integration seam — the real-but-simple isolated prototype: one artifact (script, tiny BFF, MCP server, or skill), compiling SDK snippets, constraints cited from the API docs.

Verdicts: PASS → new numbered experiment, FAIL → kill note in additional.md, RESHAPE → reframe and re-run. The full protocol — including the grilling-style question rounds and the ready-to-send reply template — is in AGENTS.md.

Running it yourself

Clone the repo and open it in any AGENTS.md-aware agent harness (opencode, Claude Code, Codex, Cursor, …). The audit loads automatically — type an idea in chat and the agent will run the gates, grill you in rounds, and write the verdict to docs/.

The audit is also packaged as an installable skill: copy .agents/skills/agent-connect-audit/ into your own .agents/skills/ (or ~/.claude/skills/) to use it outside this repo.

Bundled skills this repo uses (from Matt Pocock's skill set): grilling (the round-based interview protocol), improve and codebase-design (the aggressive quality bar). They live in .agents/skills/.

Rebuilding the site

node scripts/build-experiments-html.mjs

Regenerates docs/index.html from the experiment markdown files. Run it after any write to docs/ and confirm it exits cleanly.

Source of truth

The API's source of truth is the skills repo — start at near-intents-agent-api/SKILL.md, then the references/ files. The hosted API (https://api.agentsonintents.com) is mainnet with real funds: build with small amounts and dedicated owner accounts.