Probe

Deep research on demand

Probe lets agents and humans buy deep research: paid in USDC over x402, run by a specialist agent swarm.

Start a research run →

How it works

  1. STEP 01
    ?
    Type a question

    No login, no API key. Connect a wallet on Base when you submit.

  2. STEP 02
    ⌖
    Agents read the web

    Five sub-agents in parallel. Each one searches papers, docs, and fresh sources.

  3. STEP 03
    §
    They cite the sources

    Every claim links to where it came from. No invented references.

  4. STEP 04
    ✓
    You get an answer

    Markdown with citations. Usually back in around a minute.

Why Probe

Why use Probe

Probe is a paid research primitive for the agentic economy. Agents and builders can outsource discovery, get cited output back, and keep moving without standing up another SaaS account.

  1. CONTEXT
    Research without context burn

    Agents can call Probe for the messy web/docs/papers pass, then keep their own context for trading, building, and execution.

  2. PRICING
    Cheaper research

    Probe routes inference through Bankr's LLM gateway at zero markup. Every run settles as a real USDC x402 payment on Base - you pay for research, not platform margin.

  3. SOURCES
    Receipts for every answer

    The swarm returns sourced claims instead of vibes. Builders can inspect what was found before they route it into another agent loop.

  4. ACCESS
    Works across surfaces

    Use it from the web app, call it via MCP, or hit the x402 API directly. Pay per run - no subscriptions, no seats, no lock-in.

Pricing

Pay in USDC on Base. Your wallet signs once; the x402 facilitator settles it. No subscription.

Standard
◈
~$0.25
per query
3 agents · ~30 sec

Three specialist sub-agents fan out across web, academic, and docs sources. A synthesizer composes a cited answer. Good fit for focused questions where one good source per angle is enough.

Use standard →
RECOMMENDED
Deep
✦
~$1.99
per query
5+ agents · ~60 sec

Five-plus specialist sub-agents in parallel for harder questions. The planner picks the mix; the synthesizer cross-checks claims across more sources. Use when one angle won’t do.

Use deep →

Same question. Two answers.

What a generic chatbot tells you vs. what Probe agents actually find. Both got the same prompt:

$ ask --query
What is the current TVL of Aave v3 on Base?
✗ GENERIC CHATBOT
Generic chatbot
Clawdbot
generic LLM · no tools
Aave is a leading decentralized lending protocol that has expanded to Base, among other chains. Its TVL on Base has grown over time, though exact figures fluctuate with market conditions. I don't have access to real-time data, but based on my training data, Aave v3 on Base reached several hundred million dollars in TVL. For the current number, I'd recommend checking DeFiLlama or Aave's official dashboard directly.
✗ 0 sources cited✗ training cutoff✗ no live data✗ can't verify
✓ PROBE DEEP RESEARCH
Probe
Probe
3 agents · live web · cited
Aave v3 on Base currently holds roughly $612M in total value locked, the third-largest L2 deployment after Arbitrum and Optimism [1][2]. Borrow utilization sits around 38% with WETH and USDC as the dominant collateral assets [3].
Sources
  1. [1] DeFiLlama — Aave v3 (Base)
  2. [2] Aave governance dashboard
  3. [3] Aave docs — v3 markets
✓ 3 sources cited✓ 3 agents fanned out✓ 21 sec elapsed✓ $0.005 cost

Questions

Q 01What is Probe?
A deep-research service. You ask a question, five-plus specialist agents fan out across the web, academic papers, and documentation in parallel, and a synthesizer assembles a cited markdown answer. No accounts, no API keys - your wallet pays in USDC on Base when you submit. Built for questions a chatbot can't answer well: current data, multi-source comparisons, anything that changed since a model's training cutoff.
Q 02Why use Probe?
We're better at research than general-purpose agents. ChatGPT, Claude, generic browsing agents answer from training data plus maybe one or two web fetches - fine for trivia, weak when the answer needs current data, cross-checking, or sources you can actually click through. Probe runs a planner that fans out five-plus specialist sub-agents in parallel, then a synthesizer that's only allowed to cite what they actually found. You get an answer grounded in five fresh sources with links you can verify, not a paraphrase of stale training data with fabricated citations.
Q 03How do you stop the agents from making things up?
Sub-agents only return claims tied to URLs they actually fetched. The synthesizer is given those claims and explicitly told it has no other tools - no training-data fallback. If sub-agents return zero usable claims, the synthesizer refuses and hands back an empty answer instead of inventing one. You see the claims and sources in the output; nothing is hidden.
Q 04What do you keep about me?
Wallet address, query text, the run output, the on-chain transaction hash. That's it - no email, no identity, no tracking. We don't share or sell anything; the wallet address is the only handle you have here. Run from a fresh address if you want a clean slate per query.