Trust infrastructure for AI agents. Receipts, not vibes.
I build the full stack that answers one question: should this agent be allowed to do that? Identity, translation, audit, enforcement — every claim on this page links to a repo, a live receipt or a spec. Nothing here is a vibe.
Live: utalabs.tech · alicelabs.site · marketnow.site
Four files, one question. Each product answers it at a different layer.
| Layer | Product | What it is | Proof |
|---|---|---|---|
| 01 · translation | UTA — universal trust adapter | 8 credential formats in (JWT, W3C VC, A2A, X.509, ATC…), one canonical schema out. 12-stage verification pipeline. | x402 API — $0.01/verify, fail-closed · spec + repo |
| 02 · identity | ATC — agent trust card | Ed25519-signed capability card (ATC/1.0, RFC v3.0): who the agent is, what it may touch, who answers for it. | 57 cards live · MNR-CRL-1.0 revocation · spec + repo |
| 03 · audit | Sentinel | 10 static layers + a real L2 sandbox: every MCP server executed under gVisor, syscalls watched for 30 s. | 1,211,488 checks · 1,030 threats caught |
| 04 · enforcement | Interceptor | MCP runtime policy engine. Blocks .env reads and rm -rf even when the caller's card is valid. |
5 live rules · signed JSONL audit trail |
| annex A · provider side | UTA-Shield — firewall for MCP operators | Wraps your MCP server once: identity gates, threat-intel blocking (23 live IOCs from a real takedown), rate limits, DLP, x402 metering. 19.8 kB, zero dependencies. | repo |
| annex B · registry | MarketNow | Security infrastructure for AI agents: skills, credentials and domains verified across 71,198 MCP servers indexed. | free MCP API, 10 tools, no signup · repo · status |
Fail-closed is the house rule: unreachable revocation check → DENY. Downtime degrades to "no", never to "probably fine".
Sleepy Lazy Overnight Task Handler. A 1B open-weight brain (Gemma 3 + llama.cpp) that runs on your CPU and does the boring stuff while you sleep. Every action lands on a SHA-256 hash chain, signed with Ed25519 — offline-verifiable by anyone in ~30 seconds, no account. Irreversible clicks stay human.
- Read tonight's receipt: receipt/latest.json
- Verify the chain yourself: sloth.utalabs.tech/verify
marketnow.site · status · repo
Where the stack meets the world: security infrastructure for AI agents — skills, credentials and domains verified across 71,198 MCP servers indexed (135,547 tracked), with a free MCP API — 10 tools, no keys, no signup — plus signed submissions, revocation checks and a Sentinel score on every entry.
- Free API: marketnow.site/api/mcp
- Live status: status.marketnow.site
- Agent surface: uta-ai — installable skills, subagents, MCP servers, Claude Code plugin
Real-time professional audio measurement in the browser — RTA, spectrogram, transfer function, impulse response and SPL. The free alternative to Smaart.
- alethech — verifiable agent continuity: Ed25519-signed memory, hash-linked DAG, key rotation with reachability guarantee. Local-first, zero-LLM, zero-blockchain. Docs: alethech.alicelabs.site. Listed in Awesome-Agent-Memory after a maintainer audit of the claims against the code.
- universal-trust-adapter — ATC/1.0 protocol spec + reference impl (Ed25519, RFC 8785 JCS, 10 verification controls). One-command conformance:
npx @marketnow/uta-conformance(36 published vectors). Listed in awesome-agent-native-services after review (PR #229 merged). - marketnow-mcp — free trust layer for AI agents over MCP: verify credentials (8 formats), scam-check domains, fingerprint tools. Live: marketnow.site/api/mcp. Listed on mcp.so.
- sloptotal — open-source AI text detector: 23 engines, one calibrated score, every number measured on a public benchmark. Self-hosted, CPU-only. Live: sloptotal.com.
- mcp-suite — 229 MCP servers mapped to real agent pains: 1,035 tools, 349 functional tests.
- yunaremaia/mcp-guard — 3 merged PRs: npm attestation/provenance verification with strict policy, CLI exit-code fix on empty scans, whole-word matching for write/destructive keywords.
- run-llama/llama_index — open PR: simple fusion builds fresh wrappers instead of mutating retriever-owned ones.
- cognicore-dev/cognicore-env — open PRs: lexical-search score bounds in the SQLite memory backend, verifier-defeat chain topology tests.
- x402-foundation/x402 — spec work on the PAYMENT-REQUIRED schema: falsy-value rule re-scoped to examples with named-path failure reasons and
unavailableas a third verdict in the redraft. - taikotalent-phishing-iocs — published IoCs and attack-chain writeup for a live fake-recruiter campaign (dev.to writeup).
Long-form series on building agent trust infrastructure — what a maintainer audit actually checks, and the honest accounting of what listings pay: dev.to/edison_flores_6d2cd381b13
utalabs.tech · alicelabs.site · AliceLabs LLC — Wyoming, USA · Quito, Ecuador