langchain

LangChain Deep Agents × Circle Agent Stack

What it is

An Autonomous Payment Agent built with LangChain Deep Agents. From a single TypeScript entry point, the agent bootstraps via the Circle Agent Skill, creates an agent wallet on BASE, checks balances, discovers an x402-compatible service on the Circle Agent Marketplace, and pays for it using a USDC nanopayment.

Prerequisites

• Bun 1.2+
• An LLM provider API key (Anthropic or OpenAI)
• A provisioned agent host: Circle CLI installed, skill installed, and Circle's Terms of Use accepted. This is setup.md steps 1-2 plus the one-time Terms gate, a per-host operation. See Host setup. Login is not part of host setup because the demo logs you in with email + OTP on its first run. See Login.

Quickstart

git clone <repo-url> && cd agent-stack-ecosystem-kits
bun install
cp kits/langchain/.env.example kits/langchain/.env   # then fill in keys
bun run --cwd kits/langchain demo

Run the demo with --cwd, not bun --filter. --filter wraps output in a dashboard that elides lines and interferes with the interactive approval prompt; --cwd runs the script directly with plain, full output.

Run Host setup once before the first demo. On first run the demo logs you in with email + OTP (see Login), then pauses for your approval before any USDC payment. See Human-in-the-loop.

Environment

Variable	Required	Notes
ANTHROPIC_API_KEY or OPENAI_API_KEY	one of	Provider auto-selected from whichever key is set. Anthropic wins if both are set.
LLM_MODEL	no	Overrides the default model (claude-sonnet-4-6 / gpt-4o-mini).
NO_COLOR	no	Set to disable colored output. Color is auto-disabled when output is piped or redirected.

The kit pays on Base by default and falls back to Polygon when a service offers no Base payment option. The chain is selected automatically per service, so there is nothing to configure.

Human-in-the-loop

The agent runs the full tool loop autonomously, with two exceptions: circle_pay_service and circle_gateway_deposit, the only tools that spend USDC. Both are wired with Deep Agents' interruptOn, so the agent pauses before spending and waits for your decision. Read-only tools (skill fetch, wallet list/balance, gateway balance, service search/inspect) and circle_deploy_wallet (a zero-value, gas-abstracted wallet bootstrap that spends nothing) never pause.

interruptOn is per-tool by design. The blunter LangGraph-level interruptBefore: ["<node>"] would pause before every tool call, including read-only ones; interruptOn: { circle_pay_service: true, circle_gateway_deposit: true } pauses only on the spends. A MemorySaver checkpointer persists agent state across the pause.

When the agent pauses, the entry point reads result.__interrupt__, prints the pending tool call and arguments, and prompts for y/N approval in the terminal, the way Claude Code prompts before a sensitive action. It then resumes the agent with new Command({ resume: { decisions } }) carrying an approve or reject per pending action. Run the demo in a real terminal so the prompt is answerable.

What the demo does

The entry point passes the Circle bootstrap prompt to a LangChain Deep Agent and lets setup.md drive the flow. There is no hand-written system prompt.

0. Before the agent runs, the demo checks the CLI session. If you are not logged in, it runs the email + OTP login inline; if the Terms of Use are not accepted, it stops with the one manual step. A valid session is skipped straight through.
1. The agent calls fetch_setup_skill, reads the returned 7-step skill, and follows it.
2. Steps 1-2 (CLI install, skill install) and the Terms gate are already satisfied by Host setup; login is handled in step 0 above, so the agent picks up at wallet provisioning.
3. It lists or creates an agent wallet on BASE, checks the USDC balance, searches the Circle Agent Marketplace, inspects a service, and pays for it with a USDC nanopayment. fetch_sub_skill pulls wallet-fund / wallet-pay guidance when a step needs it.
   • A Circle agent wallet is a Smart Contract Account: its address is counterfactual until the first outbound transaction. It can receive USDC, but cannot sign x402 payments until deployed. If the paying wallet has never sent a transaction, the agent calls circle_deploy_wallet first: a one-time, zero-value self-transfer that deploys the account. circle_pay_service also pre-checks deployment (via eth_getCode) and returns an actionable error if the wallet is not yet deployed.
4. Before circle_pay_service runs, the agent pauses for approval. See Human-in-the-loop.
5. Every tool call logs to stdout ([tool] ...); the agent's reply prints after each turn.
6. The demo then drops into an interactive you> prompt. Type follow-ups ("discover services", "pay for the Bitcoin price service") and the agent keeps full context across turns. Empty input or exit / quit ends the session.

The whole run is one conversation: a MemorySaver checkpointer with a stable thread_id persists agent state across both the approval pause and every chat turn.

Skill reference

The agent boots from the official setup skill via this prompt:

Run curl -sL https://agents.circle.com/skills/setup.md, and use the returned setup instructions to set up my agent wallet.

See https://agents.circle.com/skills/setup.md.

Host setup

Run once per agent host. This is setup.md steps 1-2 (CLI install, skill install):

bun add -g @circle-fin/cli
circle skill install --tool claude-code   # or: cursor | codex | opencode | amp
# Universal fallback (any host):
bunx skills add circlefin/skills -g

Login is not part of host setup; the demo handles it. See Login.

Login

On its first run the demo checks the CLI session (circle wallet status) and, if you are not logged in, runs Circle's two-step email + OTP login inline:

1. It prompts for your Circle account email and runs circle wallet login <email> --init.
2. It prints the CLI output, which carries an anti-phishing prefix; match it against the code Circle emails you.
3. It prompts for the OTP (the 6 digits alone, or the full B1X-123456) and completes the login.

You type your own email and OTP; the kit stores neither. A first-time login also provisions an agent wallet on every supported EVM chain, so no separate circle wallet create is needed. A valid existing session is detected and skipped.

Terms of Use are not handled by the demo. If the Terms are not yet accepted on this host, login is gated and the kit stops with a manual step: run circle wallet status yourself, accept the Terms when prompted, then re-run. Per setup.md, an agent must never accept the Terms on a user's behalf, so this kit ships no Terms tool. See wallet-login.md for the full login flow.

Architecture

┌─────────────────────────────┐
│   LangChain Deep Agent      │
│   (planner + sub-agents)    │
│   interruptOn: spend tools ─┼──▶ pause ▶ human approve / reject
└──────────────┬──────────────┘
               │ tool calls
               ▼
┌─────────────────────────────┐
│  @.../circle-tools          │
│  (execFileSync → circle)    │
└──────────────┬──────────────┘
               │
               ▼
┌─────────────────────────────┐
│  Circle Agent Stack         │
│  wallets · services · x402  │
└─────────────────────────────┘

Links

• LangChain Deep Agents: docs, GitHub
• Circle Agent Stack
• Circle Agent Marketplace