Nostr Agent Onboarding · start here · index · source: Nostr-exp/core/05-current-frontier.md · snapshot 2026-10-10

Paths such as ~/Documents/…, ~/Production Environment/…, repos/… and services on localhost refer to the author's workstation and are not available to you — read them as worked examples of a setup you can recreate.

The current frontier (as of late 2026)

Part 5 of the core synthesis: where the live edge of Nostr development actually is, drawn from the freshest material in this repo — fiatjaf's 2026 notes (../voices/fiatjaf/on-nostr.md), the 2026 No Solutions episodes (../voices/no-solutions/episodes.md), and Constant's active projects (../voices/constant/research.md). Everything here is moving; re-verify before building on it.

What the principals are actually pushing right now

• Relay feeds as a first-class client feature. Both fiatjaf and Constant: a curated relay should be favorited as a feed, not added to your read list. Supported across a dozen clients (Jumble, Nostur, Yakihonne, Coracle, fiatjaf's own Hallway…). Relays with personality — communities, curation, algorithms — are the underused half of the protocol.
• Boring sub-protocols being dogfooded to reliability: NIP-17 DMs ("I just wanted proof that NIP-17 could work reliably, and now I got it" — fiatjaf), NIP-29 relay-based groups ("you only give partial authority to a server, you always keep your right to migrate"), NIP-34/GRASP git (working, still UX-rough — "will we ever get a 'merge' button?").
• The NIPs process itself is in question: fiatjaf now maintains a bare registry-of-kinds and argues for competing prose guides over numbered specs. Watch where documentation actually accrues.
• Spam doctrine consolidating around filtered read-relays (WoT-gated inbox relays like Pyramid's /inbox, NIP-05-gated, PoW, paid) with client-side WoT as a secondary layer — though Martti Malmi still runs WoT client-side on principle (trusting the relay operator is the cost he won't pay). Unresolved, productive disagreement.

Media and app distribution

• Blossom everywhere: content-addressed blobs, user-declared server lists (kind 10063), the relay design pattern applied to media. Known gap: self-healing links — clients that repair a 404 from the user's server list — are still mostly unimplemented; the No Solutions podcast itself 404'd when one Blossom server died.
• nsites: whole static sites as Nostr events + Blossom blobs (Constant's techno-ethica.com renders entirely from relays). Now specified as NIP-5A: one replaceable kind 15128 manifest per pubkey listing every file by sha256, named sites at 35128, immutable snapshots at 5128, and a 2026 addition — an aggregate hash that makes two manifests provably the same site, plus a/A lineage tags so anyone can pin and republish someone else's site under their own key. Credible exit for a website. fiatjaf's "subjective apps" sketch goes further: a web client addressed by content hash, releases adopted voluntarily through the social graph — his answer to "web apps can't be true user agents." (How to actually ship one, with the failure modes: ~/Documents/nostr-dev/docs/patterns/nsite-publishing.md.)
• Zapstore: permissionless app distribution with WoT instead of a blessed catalog; evolving from "user against the world" verification to community catalogs that do the trust heavy-lifting. The forcing function is real: OS-vendor sideloading crackdowns ("you're either a fully-KYC'd developer or you're a terrorist" — Franzap).
• Composable clients: napplets (sandboxed single-file mini-apps whose host runtime mediates signing, relays, uploads, permissions — "you should be able to compose your clients"), Soapbox Tiles, fiatjaf's revived nostr-apps idea. Direction: Nostr expertise concentrates in runtimes, relays, and libraries; app authors stop needing to know about keys at all.

Identity, trust, and time

• Signer maturity: NIP-46 bunkers, Amber/NIP-55, FROST/musig2 threshold schemes ("pomegranate"); next frontier is signing UX — specialized signers, budgets, "the user must understand what they're signing within half a second, then you can auto-approve" (hzrd149).
• Key rotation still unsolved, deliberately: the field is converging on social recovery / attestation-based migration (Constant's expiring musig2 association events; Malmi's keeper-device + per-device keys; Gigi/Pablo's "your friends tell people it's actually you") rather than any formal scheme. "You are not your npub. Your npub is a pointer."
• Provable time: OpenTimestamps integration (NIP-03 → Constant's NIP-3B/3C: stamp hash(id+sig), embed the proof) as the defense of history against key compromise and retrospective fabrication.
• Permissioned environments from permissionless blocks: Constant's TEPP / Kidstr (guardian-managed lists of people/places/things enforced at signer, client, and relay) as the freedom-tech answer to age-verification and digital-ID law — expect this whole area (child safety, SFW modes, supervised onboarding) to grow as regulation lands.

Networking beyond the social layer

• Nostr keys as network identity: FIPS (mesh routing to a pubkey instead of an IP; peer discovery and NAT traversal over relays), Malmi's Nostr VPN (WireGuard negotiated over relays, peers addressed by npub, Cashu-incentivized exit nodes), Hashtree (content-addressed filesystem + git on Blossom, convergent encryption for host deniability). The pattern: Nostr as the coordination and identity plane for arbitrary protocols — "DNS and IP are the worst offenders" of internet centralization.
• Money in-band: Cashu/NIP-60 wallets as events (your ecash follows you), nutzaps as self-verifying payment receipts, HTTP-402/bearer-token flows replacing negotiation dances, DVMs as pay-per-use compute. The restaurant model of monetization: sell something actually scarce, no compliance department.

The AI convergence

The loudest new thesis of 2025–26, and it cuts both ways:

• Nostr as agent substrate: "a neutral substrate that has money, identity, cryptography, and web-of-trust baked in" (Pablo). Agents mint their own npubs, hold NIP-60 wallets, discover tools via NIP-89/90, pay authors of signed knowledge that helped them. Signed provenance becomes existential as AI slop compounds — maintainer-signed docs and snippets outrank the model's own priors.
• AI as adoption lever: agent-assisted building collapses the cost of the micro-clients the philosophy always wanted (Malmi writes "basically zero" code by hand; Constant ships constantly as a self-described non-programmer; 80% of a SEC workshop shipped a working applet). Hodlbod's hope: AI-assisted hacking of open protocols "tears down walled gardens"; his warning, seconded by everyone: AI-built apps that skip routing, handlers, and WoT are architecturally hollow and quietly re-centralize.
• The training-data window: "we have this window of opportunity to steer the LLMs and the future of technology into a freedom-tech direction" (Malmi) — publish your thinking where models will learn it, or the next generation of tools knows only Stripe, Cloudflare, and GitHub. Early evidence it works: agent workflows "were always trying to do client-server, REST, username-password… now Nostr is in the training data" (Sandwich).

Open problems, honestly stated

1. Key rotation / compromise recovery — social schemes sketched, nothing deployed at scale.
2. Outbox-model completeness — most clients still don't fully implement discovery, migration, or re-broadcast; fiatjaf's two-notes experiment still indicts the ecosystem.
3. Blossom link-healing and content longevity (the "Nostr pacemaker" daemons) — obvious, unbuilt.
4. Bootstrapping trust for relay discovery (NIP-66 + WoT) — early.
5. Large-group private messaging — MLS (White Noise/marmot) vs double-ratchet trade-offs unresolved; group privacy at scale may be unachievable in principle.
6. Monetization for creators who need it today — "if you are starving, might as well come and starve here" (Constant). The honest pitch is pioneering, not parity.
7. The DM question — whether encrypted private communication belongs on Nostr at all (Constant: publications protocol; Malmi: no-network-effect privacy is exactly where Nostr can win; hzrd149: use Nostr as lookup, let Signal-class protocols carry the payload).