← PocketAgent · Registry

Vision & roadmap

v1 ships today. Ed25519 signing already graduated from this roadmap to live (see the “sign with my key” option in the composer and the ✓ signed badge on install). Everything below is the forward-looking thinking — kept off the main page so it stays about shipping.
§06b · landscape

One layer the agent stack hasn't claimed.

April 2026: the portable-agent stack consolidated under the Linux Foundation. MCP (donated by Anthropic Dec 2025) standardizes tool servers. A2A v1.0 (Google) standardizes agent-to-agent messaging, with Signed Agent Cards in early 2026. AGNTCY (Cisco-led, LF-hosted) standardizes identity and discovery. Claude Skills (Dec 2025, also adopted by OpenAI for Codex / ChatGPT) standardizes capability bundles. None of them standardize what PocketAgent does:

The unclaimed layer

The persona + rules + few-shot system-prompt payload, prepended to a third-party chat's system prompt by a user opt-in extension. A single-string URL fragment, not a directory of files. Different transport, different consumer, different trust model than Skills.

How v2 plugs in

v2 adopts the shape of Claude/Codex/ChatGPT Skills (name, description, plus our role, rules, fx, turns) and the signing model of A2A Signed Agent Cards (Ed25519 over a canonicalized payload). "A Skill manifest in a URL, signed by a DID."

§06c · thesis

Why this is a game changer (and how it could fail).

Steelman

Every other agent format is host-bound: GPTs live in OpenAI's store, Projects in Anthropic's account, custom instructions in the app you're typing into. Skills are the closest thing to portable, but a Skill is a directory that an agent runtime loads — not something you can paste in a DM. A URL-fragment agent works in any channel that lets you send a link. Combined with the v2 signing layer, an idea (a persona, a rule-set, a worldview) becomes a citable, verifiable, payable artifact.

Killer use case

Authors, educators, and consultants distill their own working style into a PocketAgent and share it. A book has a reading list; a researcher has a PocketAgent. Receivers install it, their everyday LLM takes on that voice, and (in v2) every chat session that uses the agent attests usage and pays the author in stablecoin. Ideas earn — the way songs earn on streaming — without a marketplace gatekeeper.

Killer objection

Browser extensions are a hard adoption surface. The active- mode injector covers ChatGPT, Claude, Gemini, Mistral — but each of those vendors could ship their own "official import" and route everyone away from a third-party fragment scheme. The defense is the open spec: PocketAgents are a format, not a service. If ChatGPT ships an import button that reads #pa=…, the format won.

What has to be true

(1) The signed identity layer (v2) ships with green-check UX, so receivers stop being scared. (2) At least one consumer client other than this site adopts pocketAgentRead() — the Skill-shape JSON makes this easy because it's already what Codex and ChatGPT understand. (3) x402 payments route to agent-author addresses by April 2027, so the "ideas earn" promise stops being theoretical. None of these require a new standards body.

§06d · roadmap

v2 — signed identity, then payments.

v1 ships today: a portable, scrubbed, spotlighted system-prompt bundle that any third-party chat with the extension can read. v2 layers provenance and payment on top, without breaking the v1 reader. A v1 client encountering a v2 payload ignores the new fields; a v2 client adds a green-check badge and attribution.

Identity (v2.0)

did:pkh:eip155:8453:0x… as canonical agent ID. Free — it's just a wallet address on Base. Optional did:web alias for human-friendly names. ENS for display, after Namechain was scrapped Feb 2026 because mainnet registration costs fell ~99%.

Signing (v2.0)

Ed25519 detached signature in a new sig field, ~88 base64url chars over a JCS-canonical payload. Browser-side pocketagent-sign.js uses WebCrypto. Verification adds <1ms. Mirrors A2A v1.0 Signed Agent Cards.

On-chain profile (v2.1)

Pin agent metadata to IPFS, anchor with an EAS attestation on Base (~$0.01 one-time). Public, indexable, no marketplace required.

Attribution (v2.2)

Consumer clients tag every LLM call with the installed agent's DID and emit an EAS attestation per session { agentDid, sessionHash, msgCount }. Off-chain (free) with on-chain UID anchoring.

Payment (v2.3)

x402 (Coinbase + Linux Foundation, launched April 2 2026) routes USDC micropayments on Base from consumer clients to author 0xSplits vaults. Stripe-Tempo MPP shipped on Base March 2026 — same rail.

Compression upgrade

Switch gzip → brotli in CompressionStream. Firefox 147 (Apr 2025), Chrome, and Safari all ship Brotli now; ~15–25% smaller for English prompt text. Keep gzip as feature-detected fallback.

Why Base, not Solana? Solana is cheaper per tx and faster to finality, and we considered it carefully. Three reasons we chose Base: (1) x402 ships Base-first, and x402 is the rail every payments incumbent — Stripe, Visa, Mastercard, AWS, Google, Cloudflare — just bet on; (2) ENS reverse-resolution costs collapsed ~99% in 2026, neutralizing Solana's cheap-name advantage; (3) viem+wagmi's TypeScript ergonomics fit scripts/pocketagent/'s zero-dependency JS aesthetic better than the Solana wallet-adapter+Anchor split. Adoption, not throughput, is the bottleneck.