The marketplace where AI agents buy verified workflow knowledge — at the moment of failure, sub-second, sub-dollar. Every block a signed unit of expertise. Every sale a 70% royalty to the creator who proved the answer once.
Most failed agent runs fail on specific, recoverable gaps — a config value, an API quirk, a workflow step. Frontier models froze before the hardware-stack combination stabilized. Your $200 Pro plan can't buy you a fix that shipped after October.
The cutoff window doesn't shrink to zero with each new model release. Every library that ships between training runs is a stuck agent waiting to happen. Stack Overflow can't help you — agents can't call SO programmatically, and the best debuggers don't write up the answer because there's nothing in it for them.
Agent hits a post-cutoff bug → MCP server returns 3 candidate blocks → agent picks best match → buys for $0.10 in USDC → injects fix → ships task. Total wall-clock time: under 1.5 seconds.
Sub-second settlement at the point of failure. No human in the middle. No subscription. No friction between a stuck agent and a shipped task. The protocol nobody used since 1997 — wired into the agent loop where it belongs.
Reproducible test cases attached. Creator signature on-chain. Price set by demand. Royalty splits automatic. When another agent resells, the original creator still earns. Knowledge that survives the company that minted it.
50 technical debugging tasks. 6 arms. 200 completions. LLM-as-judge grading. Total benchmark spend: $4.18. Context is more valuable than weights.
| Model | Block | Think | Accuracy | $/task |
|---|---|---|---|---|
| Gemma 3 · 12B | — | — | 43.0% | $0.00 |
| Gemma 3 · 12B | $0.10 | — | 90.5% | $0.10 |
| Sonnet 4.6 | — | default | 81.5% | $0.24 |
| Sonnet 4.6 | $0.10 | default | 94.0% | $0.34 |
| Opus 4.6 | — | 32K max | 81.5% | $0.05 |
| Opus 4.6 | $0.10 | 32K max | 89.5% | $0.15 |
Opus 4.6 with max extended thinking ties Sonnet's baseline at 81.5%. Maximum reasoning budget can't overcome frozen weights. A $0.10 block can — and a 12B model with one block beats Sonnet alone.
{"mcpServers": {"unblock": {"command": "npx", "args": ["@kaeva/unblock-mcp"]}}}Four tools: search_blocks, preview_block, purchase_block, inject_block. Auto-trigger on agent error-loop. Pre-fund a Coinbase Smart Wallet with $20 promo credit on signup — covers ~200 post-cutoff fixes free in beta. Works with any model that takes system-prompt injection.
Flat tier at beta. Dynamic pricing + specialist tiers come online in Q3 2026.
Every price point: 70% creator · 20% validator pool · 5% protocol · 5% burn
You already burned the three hours. You already wrote the Slack post nobody saw. Export the session, run the local SDK, mint the fix. Every time another agent hits the same wall, you earn 70% of the sale in USDC — directly, on-chain, no aggregator. Royalties pay forward forever — when a buyer resells your block to a downstream agent, you earn again.
Your agents are rediscovering the same bugs independently. Every stuck-agent hour is a Stack Overflow query that should have been a $0.10 buy. Private UNBLOCK deployments let your engineers share verified knowledge across the org — agent-to-agent, no content leaves your VPC.
Nobody has the four-way intersection UNBLOCK sits at: agent-native (MCP), post-cutoff scope, creator royalty economics, on-chain provenance. But here's who else is in the room and why they don't fill the same gap.
Free, massive, and the wrong shape. Agents can't call SO programmatically, answers aren't verified, and there's no creator economics — so the best debuggers don't contribute.
Each new model shrinks the post-cutoff window — but the window never reaches zero. Every library that ships between training runs is a UNBLOCK opportunity. UNBLOCK isn't trying to replace your model — it closes the gap the model literally cannot close alone.
MCP registries will ship. They'll host free, curated servers. UNBLOCK is a marketplace with a payment layer, creator royalties, and verified content — different mechanism, different incentive curve.
RAG retrieves what you already own. UNBLOCK transacts for knowledge you've never seen — across providers, across organizations, in 5 seconds.
I burned $25 and three hours on an MI300X / vLLM serving fix last month. The docs said it worked. Claude said it worked. Both were wrong — the fix stabilized after every frontier model's training cutoff.
That was the moment the protocol in my head became a company in git.
— Viraj Sharma, founder · Kaeva Labs
UNBLOCK is the first product of Kaeva Labs, an AI lab building agent-native infrastructure — the substrate AI agents need to remember, share, transact, and be accountable across providers.
If you're a debugger who's lost three hours to a post-cutoff bug, you're a creator. If you're running 50+ engineers on AI workflows, you're a team customer. If you're reading this and you build with conviction, write to us anyway.