Docs

What is Cloven

Other agents start at zero. Yours starts at today.

Every LLM ships frozen. Its weights stopped learning on a training-cutoff date, and everything that happened since — a depeg, a model release, a rate cut, a rug — is invisible to it. Teams paper over this with hand-rolled RAG pipelines, brittle API glue, and cron jobs that scrape and re-embed. It's the single most expensive, least differentiated part of building an agent, and everyone rebuilds it from scratch.

Cloven is that layer, built once, done right. It is an always-fresh, multi-source, citation-backed knowledge substrate that any agent plugs into for a five-minute-old picture of the world — over MCP, REST, or a one-import TypeScript SDK, paid with a prepaid USDC credit or x402 per-call (both USDC on Base, no card).

We are not another agent framework, agent platform, or agent launcher. We are the substrate every agent plugs into for fresh, attributable context.

The cold-start tax

An agent with stale context doesn't fail loudly. It fails plausibly — it answers with confidence about a world that no longer exists. That's worse than an error, because you can't see it.

The usual fixes each carry a cost:

  • Fine-tuning bakes a new cutoff into the weights and is obsolete the day it finishes.
  • RAG over your own corpus is only as fresh as your last ingest, and you own the whole pipeline forever.
  • Stitching public APIs at call time means rate limits, schema drift, no citations, and latency you pay on every turn.

Cloven absorbs that tax. One continuously-running engine ingests dozens of sources, compacts them with an LLM into a typed structured state plus an analyst-voiced brief, attaches a citation to every claim, and serves the result the instant your agent asks. You write one line. We keep the world current.

How it works

The engine is a loop, not a database dump:

  1. Ingest. Each pack pulls from its own set of sources on a fixed cadence — prices, on-chain metrics, security audits, news, social, macro feeds. Sources are isolated; one going down never takes the pack with it.
  2. Compact. A fast model distills the raw firehose into a compact, deduplicated blob, dropping noise and keeping signal. This is the cleft — Cloven cleaves context out of the flood.
  3. Narrate + cite. A stronger model writes a brief in analyst prose, and every [N] in that brief resolves to a real source URL and the exact timestamp the data was fetched. No claim is uncited.
  4. Serve. The latest state, brief, citations, and a freshness age land in Redis and stream to your agent over whichever surface you chose. New compactions push as live pulses.

The output contract is identical everywhere:

{
  "state":     { /* typed JSON, validated by the pack's Zod schema */ },
  "brief":     "Analyst prose with inline [1] citations …",
  "citations": [{ "n": 1, "url": "https://…", "fetchedAt": "2026-…" }],
  "freshness": { "generatedAt": "2026-…", "ageSeconds": 84 }
}

freshness.ageSeconds is your contract. If a read is older than your tolerance, hard-fail, fall back, or retry — Cloven never invents data. If a source is down, the pack returns the last good blob with an honest age, never a hallucination that sounds current.

Three surfaces, one contract

Pick the one that matches how your agent is deployed. They return the same shape.

  • MCP server — the primary surface. Drop Cloven into claude_desktop_config.json, Cursor, Windsurf, or Zed and the cloven.fresh, cloven.brief, cloven.search, cloven.snapshot, cloven.subscribe, and cloven.cite tools appear in the model's tool list. Stdio for local clients, HTTP for remote.
  • REST — GET /v1/fresh?pack=crypto with a Bearer header. Plain HTTP, no client library, JSON back.
  • SDK — pnpm add @cloven/sdk, construct once, call .pack("crypto").fresh(). ESM-only, under 50KB gzipped, handles x402 retries automatically when you pass a wallet.

See the Quickstart to ship a working call in five minutes.

Packs — verticals, not one monolith

A pack is an isolated module with its own sources, compactor, briefer, and schema. The engine is domain-agnostic; packs are how you scope it.

  • Crypto (live) — CoinGecko, DefiLlama, GeckoTerminal, DexScreener, GoPlus security, CryptoCompare news, Fear & Greed, GDELT, Reddit. Prices, TVL, pool data, rug flags, and a market narrative, refreshed every five minutes.
  • AI (shipping) — Hugging Face trending, GitHub trending, ArXiv, Hacker News, r/MachineLearning.
  • Markets (shipping) — SEC EDGAR, FRED, earnings feeds, macro news.
  • Custom (roadmap) — bring your own sources via webhook or scraping config.

Full detail in the pack catalog.

Pricing you can start on for free

100 calls a day, free, no card — just claim your session with a magic link. Past that, two rails, one treasury, both USDC on Base:

  • Prepaid credits for humans — deposit USDC, get credits, each call debits one. Volume tiers take the per-call price from $0.001 down to $0.0006.
  • x402 per-call for autonomous agents — no key, no signup. Your agent pays USDC on Base per request via the X-Payment header. This is the same standard Coinbase, AWS, Visa, and the Anthropic foundation are building on — real institutional rails, not a memecoin.

The full tier table and per-call rates live on the pricing page.

Where crypto fits

Crypto is Cloven's first vertical pack, not its identity. The web3 surface is deliberately narrow and real: x402 native pay-per-call, optional on-chain attestation of a brief's provenance, and — later — a stake-tier discount for power users. Working rails, no LARP.

The roadmap, briefly

Cloven ships in phases, and the substrate is the whole point of phase one:

  • Phase 1 — Cloven Mind (now) — the context layer. One engine, live packs, three surfaces, on-chain billing, a console.
  • Phase 2 — Cloven Agent — deploy a curated, persona-driven agent powered by the packs you subscribe to, posting to X, Telegram, or Discord on a schedule.
  • Phase 3 — Cloven Commons — the reasoning traces accumulated from day one become a citation-attributed dataset for the labs training the next generation of agentic models, with revenue shared back to operators.

Each phase compounds on the data the last one produced. The fresh context you read today is the training signal someone pays for tomorrow.

Start here