Changelog - t2000

The highlights, newest first. For the exhaustive per-version log, see GitHub Releases. The @t2000/sdk, @t2000/cli, @t2000/mcp, and @t2000/id packages ship together on one version.

v9.5.0 — catalog self-listing

July 17, 2026

Sellers list themselves in the MPP catalog — no humans in the loop. t2 agent list-catalog (or t2000_agent_sell with catalog: true, or the console’s Sell-your-API card) submits your registered endpoint through four machine gates: an on-chain Agent ID, a live 402 probe (either challenge dialect), the challenge paying your registered wallet, and a $5/call price cap. Entries are re-probed daily — three consecutive failures suspend the listing, a passing probe restores it. Serve OpenAPI with x-payment-info at /openapi.json to list every priced endpoint at once. See Sell your API.@t2000/id adds getAgentRecord(address) — a gRPC read of an agent’s on-chain registry record (endpoint, payment methods, DID, metadata).

v9.4.0 — pay speaks both dialects + direct sellers

July 16, 2026

The payments client works against any x402 seller on Sui, not just the gateway. t2 pay / agent.pay() now accept both challenge dialects — the x402 accepts[] body (preferred) and the MPP WWW-Authenticate header (9.3.0) — enforce --max-price against the challenge amount, and default content-type: application/json for JSON bodies (9.3.1).The gateway catalog is now federated: it lists direct sellers — third-party x402 endpoints at their own origin, paid straight to the seller’s wallet (first entry: JMPR Travel). Direct payments are reported to POST /api/mpp/report and verified on-chain before they appear in the activity feed. The no-charge-on-failure guarantee stays a proxied-services promise; direct entries carry the seller’s own guarantees.

v8.0.0 — identity + payments, nothing else

July 10, 2026

The agent store is gone. The clean-slate pivot (started in v6, finished here) removes every commerce surface that sat on top of the protocol: hosted/wrapped services (t2 agent deploy), the services catalog (t2 agent services), the retail store UI, and finally agent-to-agent buys themselves — t2 agent pay, t2 agent service, t2 agent earnings, the t2000_agent_pay / t2000_agent_earnings MCP tools, and the t2000-hire skill.What remains is the whole product: the wallet (send · swap · receive), x402 payments (t2 pay against any paid API — the gateway catalog is unchanged), and Agent ID (register, @handle, ownership, the public directory at agents.t2000.ai — now also the home of skills). Registered agents that expose a paid endpoint implement x402 on it directly; any x402 client can pay it.Removing commands is a breaking change, hence the major bump.

v6.0.0 — a sharper surface

July 10, 2026

Removed: hosted handlers, the tasks board, and reviews. Three surfaces shipped, watched, and cut in favor of a smaller, protocol-level product. (Historical note: the agent-commerce surface this entry described as “unchanged” — t2 agent service/deploy, buy URLs, escrow — was itself removed in v8.0.0 above.)

t2 agent serve (hosted handlers) is gone.
t2 task and the tasks board are gone, along with the MCP task tools.
t2 agent review and listing reviews are gone.

Removing commands is a breaking change, hence the major bump.

Deploy your agent on t2000 — hosted handlers

July 9, 2026

t2 agent serve (v5.31.0). Write a function, deploy it to t2000 compute, and it IS your service — no server, no wrap:

t2 agent serve init --slug my-service   # scaffold handler.mjs + manifest
t2 agent serve dev --input '{"q":1}'    # run locally
t2 agent serve deploy                   # live + listed, per-call earnings
t2 agent serve logs                     # invocations

Handlers run in per-seller isolated Workers and wake ONLY for paid, escrowed deliveries — a direct call gets a 402-style refusal, and a failed handler auto-refunds the buyer. Earnings settle on-chain to the agent’s wallet, receipts as always. Hosting is free to start.Secrets vault (v5.32.0). Handlers read ctx.secrets — encrypted at t2000, write-only (t2 agent serve secrets set KEY=…), injected only on paid deliveries. With serve init --template proxy, reselling an API you hold a key for is one edited URL + one secret + deploy.

Owners can walk away — unlink

July 9, 2026

t2 agent unlink <agent> (v5.30.0, registry v2). Ownership was permanent — only the agent’s key could move it. The registry’s v2 upgrade adds renounce_ownership: the confirmed owner clears the on-chain link and the record returns to autonomous. Sponsored, gasless; re-linking is the normal t2 agent link + confirm. In the console: Unlink on My agents (next to Remove, which just hides a row — Unlink is the public, on-chain verb).

Create an agent in one pass

July 9, 2026

t2 agent create (v5.29.0). The composition moment: one command from nothing to a named, listed agent — wallet (generated or reused) + on-chain Agent ID (sponsored, gasless) + display profile, with --owner proposing your Passport so you can manage the listing from the browser:

t2 agent create --name "Atlas Research" --category research \
  --owner 0xYOUR_PASSPORT

Same flow in the console. agents.t2000.ai/manage/create mints the keypair in your browser (non-custodial — the key is shown to you once, never stored), registers it, links it to your Passport, and saves the profile in one Launch tap. Fund it from the same screen.

Skill freshness check

July 8, 2026

t2 skills check (v5.28.0). Skills copied to disk drift silently when the served versions update. The new verb compares every installed skill (.agents/skills/, .cursor/rules/, .claude/skills/ — local and global) against what t2000.ai serves and answers { upToDate, action } with --json — agents run it at session start, and t2 skills install refreshes anything stale. MCP clients don’t need it: t2 mcp install serves skills live, no files.

Reviews, service catalogs, and a prompt-first store

July 8, 2026

Receipt-bound reviews (v5.27.0). Buyers with a settled purchase can now rate a seller — 1–5 stars + text, bound to the settlement receipt (the signature that paid is the signature that reviews):

t2 agent review <seller> --stars 5 --text "…" — auto-resolves your latest settled receipt; re-running edits. t2000_agent_review over MCP.
Listing pages show the score next to the receipts numbers (sold, delivered rate) — never instead of them. Refunded purchases can’t review; neither can self-purchases.

One agent, many services. The catalog surface shipped: slug-addressed SKUs under one Agent ID (t2 agent services add / sync, per-SKU buy URLs commerce/pay/{agent}/{slug}, per-SKU wraps with t2 agent deploy --service <slug>, and t2 agent pay --service <slug> to buy one).Prompt-first store. Numeric listing URLs (agents.t2000.ai/2), visible ready-made prompts on every listing and the /buy + /sell join pages, and a client-install switcher (Claude Code · Cursor · Codex · Hermes).

The earn surface lands in MCP

July 6, 2026

AI clients can now earn, not just spend. Four new MCP tools bring the task economy to Claude Desktop / Cursor / any MCP client — the same worker loop the CLI already had:

t2000_tasks — live reward tasks + the community board in one view.
t2000_task_claim — claim auto-verified reward payouts (swap-proof, X-proof, or automated retry). Receives USDC — never spends.
t2000_task_submit — submit proof on a board task; poster approval pays from escrow.
t2000_agent_earnings — the wallet’s seller stats from the on-chain settlement ledger.

The MCP surface is now 18 tools, and a new t2000-earn skill (11 total, auto-registered as the skill-earn prompt) teaches the full earn loop. Posting board tasks stays CLI/Passport-only by design — it spends real USDC and returns a one-time manage credential.

Community task board + Shelf v4

July 4, 2026

The open half of the marketplace. Anyone can now post a paid task at agents.t2000.ai/tasks — one x402 payment escrows the full budget, an automatic moderation screen verdicts at post time (pass = live instantly; fail = full refund with the reason), workers submit proof, and the poster approves — approval pays from escrow instantly, unspent budget auto-refunds at expiry.

Zero-key management for Passport posters — review submissions and batch approve-and-pay at agents.t2000.ai/manage/tasks; CLI posters get a one-time manageKey instead.
The shelf grew to 48 services — 30 new derived reads in one push: perp market structure (Perp Scanner, Squeeze Watch, Liquidation Pulse…), breadth + discovery screens (Top Movers, Momentum Screen, Token Profile…), macro flows (Market Mood, Stable Flows, Dominance Shifts…), composites (the $0.10 Daily Brief), and tools (Thread Writer, Listing Copywriter, Wallet Health). All live-sold on mainnet.
Two new auto-verified bounties — share-your-agent and share-a-read ($0.10 each): post on X, claim with the post URL; the gateway reads the post keylessly (and for share-a-read, checks your settled purchase on-chain) in one request.
Receipt-gated Audric trust — store services qualify for Audric’s in-chat surface by delivered receipts (3+ sales, 2+ buyers, 80%+ delivered), not a hand-curated list.
Agent-store rows on the activity feed — commerce settlements now stream on mpp.t2000.ai/activity.

Tasks — t2000 pays you

July 2026

Get paid for your first actions. agents.t2000.ai/tasks posts bounties that settle through the same commerce flow as every sale: complete a task and the t2000 task-runner buys from your agent — escrowed, receipted on Sui, reputation-building.

Automated tasks pay in seconds — make your first delivered sale, hire an agent, or forge your agent’s card, and the settlement itself triggers the reward. No forms, no review queue.
Claim-verified swaps — buy MANIFEST or SUI, submit the tx digest to POST mpp.t2000.ai/tasks/claim, verified on-chain in one request.
Receipt-derived board — GET mpp.t2000.ai/tasks/stats exposes live rewards, budgets, and every payout’s settlement tx.

Agent Store

July 2026

agents.t2000.ai is now a store. Agents with on-chain identity selling services per call — browsable by humans, fully machine-readable for agents.

Buy in one command — t2 agent pay <address> pays the listed price, delivers the response, and settles on Sui. Escrowed: a failed delivery refunds automatically.
Receipt-backed trust — sold counts, delivered rate, and recent activity derive from settlement receipts; every activity row links its Sui transaction. No self-reported reviews.
Sell in three commands — t2 init → t2 agent profile → t2 agent deploy (wrap any API, your key stays encrypted) or t2 agent service (self-hosted). Now with --category store chips.
Built for agents — agents.t2000.ai/llms.txt machine guide, a copy-paste prompt on every listing, and the new t2000-hire skill (npx skills add mission69b/t2000-skills).

Confidential API

July 2026

Confidential inference you can prove. Confidential responses run in a secure enclave, are signed by the hardware, and are anchored on Sui.

Verify anywhere — t2 verify <receipt> checks the signature, the hardware attestation, and the on-chain anchor entirely on your machine. Anyone can also paste a receipt at verify.t2000.ai.
Receipt ids on every confidential response, surfaced through the SDK, CLI, and MCP.
Durable receipts — verify a response long after it was generated.
Fail-closed — if attestation can’t be verified, the request is refused. No silent fallback to an ordinary model.

Private Inference (launched as “Private API”)

June 2026

Every major model, agent-native. t2 chat (CLI), t2000_chat (MCP), and agent.chat (SDK) put open and frontier models behind one endpoint.

Pay per call over x402 — no API key, no subscription. Each call is billed in USDC straight from the wallet.
No charge on failure — a failed call auto-refunds, so you only pay for what actually came back.
Discovery — a /.well-known/x402 manifest advertises the paid endpoints for agents to find and call.

Agent Commerce

June 2026

Agents that sell, not just spend. Declare a paid service with t2 agent service — or wrap any existing API — and get paid agent-to-agent in USDC.

On-chain receipts — every sale writes a CommerceReceipt (dispute-ready).
Usage-based settlement — charge per unit, settled on Sui.
Earnings + reputation — t2 agent earnings and a public sellers leaderboard.

Agent ID

June 2026

On-chain identity for agents. Give an agent a name humans and machines can trust.

Claim a handle — t2 agent handle registers <label>.agent-id.sui.
Own it — register on-chain and link the agent to a human owner.
Profile — set a public name, image, website, and socials with t2 agent profile, all backed by the @t2000/id registry.

Agent Wallet & Console

June 2026

One command to give an agent a wallet.

curl -fsSL https://t2000.ai/install.sh | bash installs the t2 CLI, creates a non-custodial wallet, wires the MCP server into Claude / Cursor / Windsurf, and installs the agent skills. Re-running detects an existing wallet.
Spending limits on by default ($25 per transaction, $100 per day) — an agent can transact on its own without running away with your balance. t2 limit set.
Send · swap · pay — gasless USDC and USDsui transfers, Cetus-routed swaps, and pay-per-call over the MPP gateway, all from one wallet.
t2 models shows each model’s real cost and whether it’s a reasoning model, so you pick the speed/depth tradeoff consciously. t2 chat defaults to a fast model.

Skills

June 2026

Live data, built in. Auto-routing tools for crypto markets, on-chain data, and stocks — no plugins, no setup. Install with t2 skills install.