Agentic Finance Graph

Home/Research/MCP server

MCP server · free · read-only · no key

Ask the ledger from your AI

One URL, twenty-five read-only tools and five ready-made prompts: the same numbers and definitions the site publishes, and a check before your agent pays anyone.

Agentic Finance Graph · Published · Live since 28 September 2026 · Updated 30 October 2026

MCP server · free · read-only Connect your AI to the ledger: add https://agenticfinancegraph.com/mcp as a remote MCP server and ask it, in plain language, which AI agents on Base actually pay, how much of their money was really spent, what any agent has done, what our detectors saw, and what is known about an agent or an address before yours pays it. No key, no account, nothing that can move funds.

Connect it in one step

The endpoint speaks the Model Context Protocol over Streamable HTTP. It needs no authentication and keeps no session: every request is answered on its own.

1Claudethe desktop app and claude.aiEasiest

  1. Open Settings, then Connectors.
  2. Choose Add custom connector.
  3. Paste this address and save. Then say the sentence under these cards.
https://agenticfinancegraph.com/mcp

2Claude Codein your terminal

  1. Run this once.
claude mcp add --transport http afg https://agenticfinancegraph.com/mcp

3Cursorfile .cursor/mcp.json

  1. Put this in the file and restart Cursor.
{
  "mcpServers": {
    "afg": {
      "url": "https://agenticfinancegraph.com/mcp"
    }
  }
}

4VS Codefile .vscode/mcp.json

  1. Put this in the file; VS Code offers to start it.
{
  "servers": {
    "afg": {
      "type": "http",
      "url": "https://agenticfinancegraph.com/mcp"
    }
  }
}

5Apps that start a local programfor example Claude Desktop's config file

  1. Some apps can only run a small program on your computer (MCP calls this "stdio"). Add this; it passes every message on to the address above. Needs Node 18 or newer.
{
  "mcpServers": {
    "afg": {
      "command": "npx",
      "args": ["-y", "github:AgenticFinanceGraph/agentic-finance-graph-mcp", "mcp"]
    }
  }
}

6Your own agentany framework with an MCP client

  1. Point its MCP client at the address in card 1. No key, no account.

No AI at all? Ask from a terminal:

npx -y github:AgenticFinanceGraph/agentic-finance-graph-mcp find 57657

Then just say this

No commands to learn. Once it is connected, tell your AI in your own words, for example:

I connected Agentic Finance Graph (AFG). Look through what it offers and show me what we can do with it.

The server tells your AI what it holds, so it reads today's figures and suggests what to ask next: an agent's record, a check before paying one, where agent money goes. In Claude you can also pick start_here from the prompt menu.

Source and CLI on GitHubOfficial MCP RegistryGlama

Apache-2.0, no dependencies. In the official registry as com.agenticfinancegraph/agentic-finance-graph, with the publisher verified by our domain.

Works with your wallet

The ledger holds no keys and moves no money; your agent's wallet does. These recipes put a wallet an agent can pay from and this server in the same client, so the agent can check the other side before it pays. Every command was read from the wallet's own documentation or package on the date given. Nothing here is a partnership or an endorsement, and nothing is claimed about those products beyond what their pages say. Our side is the same each time: the address in card 1, the stdio bridge in card 5, or the afg command line.

The rule to give your agent, in its instructions:

Before any payment to an AI agent or to an address you were given, call pay_check (for an agent) or payee_facts (for an address) on Agentic Finance Graph. If refuse_conditions_met is not empty, do not pay; tell me which condition and show the evidence. If the agent is not ranked, say so: unknown is not bad, but it has no payment history we can see.

ACoinbase Agentic Walletpayments MCP · read 30 Oct 2026

  1. Install the wallet's MCP (Node 22 or newer). The installer configures Claude Desktop, Claude Code, Codex CLI or Gemini CLI.
  2. Ask the agent “Show me my wallet”, verify your email in the browser window that opens, fund it with USDC. Its FAQ names USDC on Base, Polygon and Solana, and limits per call and per session that you set in the wallet's own UI.
  3. Add this server (card 1 or 2) and give the rule above.
npx @coinbase/payments-mcp

BAgentCashnpm agentcash 0.17.1 · read 30 Oct 2026

  1. One line in Claude Code (or npx agentcash install for a guided setup; codex mcp add agentcash -- npx -y agentcash@latest in Codex). A wallet is created at ~/.agentcash/wallet.json; deposit USDC on Base before paid calls. Its README lists Base, Ethereum, Optimism, Arbitrum and Polygon.
  2. Add this server and give the rule.
claude mcp add agentcash --scope user -- npx -y agentcash@latest

CCircle Agent WalletsCircle CLI 1.2.2 · read 30 Oct 2026

  1. Install the CLI (Node 20.18.2 or newer) and log in with your email. Wallets are created on every chain Circle supports, Base among them (identifier BASE; circle blockchain list prints the current set).
  2. List the agent wallet and fund it from a wallet you own (circle wallet fund --address 0x… --chain BASE --amount 10 --method crypto).
  3. The agent pays with the CLI; before it does, it asks this server (card 6) or runs afg facts <address>. Circle's own MCP at api.circle.com/v1/codegen/mcp is for building with their SDKs, not for paying.
npm install -g @circle-fin/cli
circle wallet login [email protected]
circle wallet list --type agent --chain BASE

DAlchemy CLI agent walletAlchemy's guide of 29 Jul 2026 · read 30 Oct 2026

  1. Install the CLI (Node 22 or newer), sign in, connect a session: the browser opens and you approve what the agent may do and for how long. Alchemy's guide says a session limits capabilities and duration, not amounts.
  2. The agent reads the command reference with alchemy agent-prompt. Give it the rule, with our command line beside it: afg facts <address> before a send.
  3. End the session with alchemy wallet disconnect.
npm i -g @alchemy/cli@latest
alchemy auth
alchemy wallet connect --mode session

EFireblocks x402 agentMCP server from their repository · read 30 Oct 2026

  1. Clone fireblocks/x402-agent, install, copy .env.example to .env with your Fireblocks API key, vault account id and private-key path, build. It runs over stdio (node dist/mcp-server.js, Node 18 or newer).
  2. Limits live in the Fireblocks Policy Engine; the agent also honours ALLOWED_MERCHANT_HOSTS and REQUIRE_CONFIRM. It signs EIP-3009 and Permit2; their example is USDC on Base.
  3. Add this server as a second entry and give the rule.
git clone https://github.com/fireblocks/x402-agent
cd x402-agent && npm install && cp .env.example .env && npm run build

FTurnkey signer + an x402 client@turnkey/viem 0.14, @x402/fetch 2.29 · read 30 Oct 2026

  1. Turnkey has no payment MCP. The agent is a delegated user with no permissions until a policy grants them, and Turnkey's policy engine checks every signature inside the enclave. createAccount from @turnkey/viem gives a viem account; x402's client registers it as the EVM signer (new ExactEvmScheme(account)).
  2. Put our gate hook in front of the signature: before the scheme signs, it asks gate_check and aborts on the decisions you choose. The policy engine still has the last word. Source: bin/x402-gate-hook.mjs in our repository.
import { makeGateHook } from "agentic-finance-graph-mcp/bin/x402-gate-hook.mjs";
client.onBeforePaymentCreation(makeGateHook({ agent: "57657", policy_name: "recommended-v3", abortOn: ["block"] }));

GAmazon Bedrock AgentCore GatewayAWS developer guide · read 30 Oct 2026 · not yet run by us

  1. AgentCore Gateway takes an MCP server as a target with no outbound authorisation, the way AWS documents the Coinbase Bazaar. Add this server as a target; the agent sees its tools through the gateway's /mcp URL.
  2. AgentCore Payments (in preview) settles x402 when a tool answers 402, so the same agent can buy our evidence packs. We have not run this ourselves; the commands are AWS's.
agentcore add gateway-target --gateway MyGateway --name afg --endpoint https://agenticfinancegraph.com/mcp
agentcore deploy

A wallet missing here is not a wallet that does not work: anything with an MCP client or an x402 client can do the same. Tell us what you run and we will read its documentation: contact.

Ready-made prompts

The server also offers five prompts. In Claude they appear in the prompt menu once the connector is added; pick one and fill in an agent id where asked.

  • Start here. A guided tour with today's real figures: what the ledger measures and the questions worth asking it.
  • Daily brief. What moved in the last 24 hours: rank, spending and new detections, each figure with its time and definition.
  • Check before paying. An agent's statement, level and flags, and who has paid the address you are about to pay. It refuses on a stale statement, a broken chain or open flags.
  • Explain an agent. One agent in plain words: its level and why, what it paid and to whom, and what the data cannot show.
  • Where agents pay. The endpoints the most ranked agents pay, with counts.

What you can ask

Type any of these into a connected client. Each one was run against the live server on 11 or 12 October 2026 (the last three on 30 October 2026), and the tool your AI calls is shown under it. Put your own agent ids and addresses where ours are.

The market01

“How much money did AI agents on Base really spend, and how much of it left their control?”

market_state

The market02

“How much passed through agent wallets without being spent, and why is it left out of spending?”

market_statedefinition

The market03

“How many agents are at L7, L8 and L9 right now, and what does each level mean?”

market_stateranked_agents

Other chains04

“How many ERC-8004 agents are registered across every chain, and what did their owners pay on BNB Smart Chain in 30 days?”

market_state

One agent05

“Look up agent 57657: is it ranked, what is its wallet holding, and who does it pay?”

find_agentagent_record

One agent06

“Is wallet 0x… bound to a registered agent, and at what level?”

find_agent

The rank07

“Show the top ten agents at L9, the ones paying three or more counterparties a month.”

ranked_agents

Before you pay08

“Run a pay check on agent 57657: is its statement fresh, any live incident, and has it paid anyone since?”

pay_check

Before you pay09

“Show agent 57657's latest 3-hour statement and check its proof before I pay it.”

agent_statement

Before you pay10

“Who has paid the address 0x…, and how many of those payers are ranked agents?”

counterparty_preview

Reputation11

“Who has paid agent 54261, and how many of its reviews come from wallets that paid it first?”

payer_set

Where agents pay12

“Which endpoints do the most ranked agents pay?”

counterparty_preview

Evidence13

“What would a payment-graph evidence pack on agent 57657 cost, and how do I buy it?”

evidence_pack_quote

Trends14

“Did counted spending move materially in the last 7 days?”

since_last

Trends15

“Which metrics do you measure about bridges, and from when is each one comparable?”

list_metrics

Safety16

“How many agent wallets were targeted by address poisoning this week, how many lost money, and how has it moved over 30 days?”

market_statemetric_history

Safety17

“Show the detections of the last week and explain the drain rule.”

detections

Definitions18

“What exactly counts as spending, and what does that figure not claim?”

definition

Before you pay19

“Give me the measured facts about the address 0x… before my agent pays it, and run the gate check for $5 under recommended-v3.”

payee_factsgate_check

Before you pay20

“Does agent 57657 hold a signed Agent Pass right now, and what did it check?”

agent_pass

Safety21

“What would a stolen key move if agent 57657 were compromised at this block?”

blast_radius

The tools

ToolWhat it returns
market_stateThe headline figures: counted spending, daily spending, what went into positions the payer still holds, what left its control, the routing hops and older history excluded, agents at L7, L8 and L9, concentration, detections, poisoning and the other chains. Each with its value, unit, measurement time and definition id, and the headline set beside its previous stored sample.
find_agentAn agent by token id, slug, wallet or declared name, or up to 25 ids at once: whether it is ranked, its level and when it was set, its age, bound wallet, stablecoin float, payment totals and open incidents.
agent_recordThe free record of a ranked agent with its 25 newest payments and counterparties.
ranked_agentsThe top ranked agents, optionally from a minimum level (7, 8 or 9).
detectionsThe newest detections, filtered by agent, kind, severity, time or live incidents only.
definitionThe frozen definition of any metric: method, exclusions and what it does not claim.
metric_historyOne metric over 24 hours, 7, 30 or 90 days, with its own definition, the dated changes of ours inside the range, and the date from which it compares. Gaps are real gaps.
list_metricsEvery metric we measure: key, unit, definition, first stored sample and the date from which it compares.
since_lastOne figure beside the stored sample it moved from: the previous sample, or the newest one at least 24 hours, 7 days or 30 days older. No prior sample is a gap, never a zero change.
counterparty_previewWho paid a receiving address: counted payments, distinct payers and distinct L7+ payers, first and last seen. Counts only. Without an address, the endpoints the most ranked agents pay.
agent_statementAn agent's latest 3-hour statement with its hash, the hash it chains to and a Merkle proof against the window's signed root. Check it before you pay.
pay_checkOne free call before paying an agent: statement freshness and chain, live incidents, detections in the last 72 hours, the token approvals its wallet has given, last payment, payments since the statement, and which refuse conditions are met. Facts, never a verdict.
payer_setThe wallets observed paying one agent, and how much of its ERC-8004 feedback came from them before they rated it. The list is ready to pass to getSummary as clientAddresses.
agent_stabilityThe behaviour clock of a ranked agent: complete days since its own payment behaviour last broke (a CUSUM on payments, value, payees and new payees), and whether it runs on a schedule or in bursts.
agent_independenceWhether an agent's rank rests on demand from outside its own operator set: the score, the operator sets behind the rank and the addresses that carry it.
ladder_dynamicsHow ranked agents end: the cumulative incidence of drained, abandoned, dormant and none at 30, 60, 90 and 180 days with the numbers at risk, and the launch platforms that drain more than the market.
verify_sum_logRecomputes one agent's proof against the signed head of the Merkle-sum log: the leaf, the path, the total it commits to, and the Ed25519 signature. Trust nothing you did not recompute.
agent_patternThe out-of-character score of an agent's newest counted payment against its own earlier payments (behaviour.out_of_character.v3 since 7 Nov 2026): none, notable or major, with every reason. Advisory: nothing is blocked from it.
payee_factsWhat the chain and our ledger say about an address an agent is about to pay (payee.facts.v2): its code kind and delegation, who has paid it and when, incident history, the usual size of payments into it; each fact with its source, or marked unavailable.
gate_checkOne call before a payment, in the AP2 constraint vocabulary: the payee facts, the agent's own context and the policy's tests, answered block, warn, log or allow with every reason. Advisory: the client decides, we never block.
agent_passThe Agent Pass: eight checks on one registration at its last evaluation, valid three hours, Ed25519-signed (key at /.well-known/afg-pass-keys.json). What held, what failed, what could not be checked. Holds or not, never safe.
blast_radiusWhat a stolen key would move at one block: the priced stablecoin and token balances and ETH of an agent's bound wallets, and the EIP-7702 delegate if any with its code hash.
canary_registerRegisters the public part of a decoy you planted (an address, the hash of a key or URL, a tool name or an x402 resource) so the watcher reads it. The secret never leaves you.
canary_statusEvery canary you registered with the watcher's last reading and its newest events: what touched the decoy, when, with what evidence.
evidence_pack_quoteThe price and address of a paid evidence pack. It quotes; it never pays.

What is free and what costs money

Everything the MCP server returns is free. Dated evidence packs about one agent (liveness, payment graph, drain check, incidents, peers, a bundle, a live re-probe, its statement history) or about one receiving address (who paid it) cost 1 to 30 cents in USDC on Base and are bought over x402 at /check/<kind>: the endpoint answers 402 with the exact price, and your client retries with a signed payment. A Desk plan includes every pack. All prices →

For machines

  • Endpoint: POST https://agenticfinancegraph.com/mcp (JSON-RPC 2.0; protocol versions 2025-06-18, 2025-03-26, 2024-11-05; JSON responses, no SSE stream).
  • The same data as JSON: /api/state, /api/agents, /api/detections, /api/def, OpenAPI, llms.txt.
  • Every figure carries a definition id: call definition before quoting a number, and quote the time it was measured.

Quickstart: check an agent before your agent pays it (5 minutes, free)

  1. Connect. Add https://agenticfinancegraph.com/mcp to your client (Streamable HTTP, no key, read-only). In Claude: Settings → Connectors → Add custom connector → paste the URL.
  2. Call pay_check from any shell:
    curl -s -X POST https://agenticfinancegraph.com/mcp \
      -H "content-type: application/json" -H "accept: application/json, text/event-stream" \
      -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"pay_check","arguments":{"agent":"57657"}}}'
  3. Read the answer. level is the rung on the ladder (L7 paid someone, L8 also funded, L9 repeat independent counterparties); open_incidents lists live harm-shaped detections; statement is the latest signed three-hour statement (age_hours, stale, chain_ok, flags); detections_72h and token_approvals show recent detections and unlimited approvals. The decision stays with your client: we report facts and which refuse conditions are met, never "safe".
  4. Common answers. An unranked agent has no record: we have not seen it pay, which is not an accusation. A stale statement means the agent's window is not published yet; read age_hours. GET /mcp answers 405 because MCP needs POST.
  5. Next. gate_check applies your own payment policy (AP2 vocabulary) and answers block, warn, log or allow with every reason; payee_facts checks the address you were given.

Questions people actually ask

Does the MCP server cost anything?
No. It is free, needs no key and keeps no session. Only the dated evidence packs cost money (1 to 30 cents in USDC over x402), and the MCP server quotes them without buying.
Can the MCP server move funds or sign anything?
No. Every tool is read-only. Agentic Finance Graph never holds keys, moves funds or sits in the payment path.
Which chain does it cover?
Payments are measured on Base (eip155:8453) in USDC, USDT, EURC and DAI. Registrations are counted on Base, with other chains in the census.
How fresh are the answers?
Every figure carries the time it was measured. Wallets are swept every three hours, some ranked wallets in real time, and the tape is sampled continuously.
Where is the source code?
On GitHub at github.com/AgenticFinanceGraph/agentic-finance-graph-mcp, under the Apache-2.0 licence: the stdio bridge, the command-line tool and the tool list. It has no dependencies, and every answer still comes from the same read-only endpoint.
Is it in the official MCP Registry?
Yes, as com.agenticfinancegraph/agentic-finance-graph, with the publisher verified by our domain. Glama lists it too.
Which wallets does it work with?
Any wallet an agent can pay from, placed next to this server in the same client. The page gives checked recipes for Coinbase's Agentic Wallet, AgentCash, Circle Agent Wallets, Alchemy's CLI wallet, Fireblocks' x402 agent, a Turnkey signer with an x402 client, and Amazon Bedrock AgentCore Gateway. The ledger never holds keys; it answers questions before and after a payment.