paybond-kit

Your AI agent shouldn't accidentally spend $20,000.

An unsupervised agent that can book travel, call paid APIs, or buy compute can also loop through “approved” tool calls until the budget is gone — each charge succeeds, and nothing proves the work was delivered. Per-transaction caps alone do not stop that.

paybond-kit is the Python SDK for delegated agent spend controls: authorize against a bounded budget and policy before a paid tool runs, then release or refund only after signed completion evidence. Wrap three lines around any paid tool call — authorize → execute → verify. Keep your runtime. Keep your payment rails (prefer Stripe Connect / x402; Stripe ACH Direct Debit and USDC-on-Base are also available). You get spend authorization, evidence-gated settlement, refunds, disputes, and a portable signed receipt — without building your own proof-gated release middleware.

Under the hood, Kit opens hosted Gateway sessions, verifies capability tokens, signs intent and evidence payloads, and can read tenant-scoped Signal, fraud, ledger, protocol, and A2A data, with adapters for common agent runtimes.

Docs · Python quickstart · Support matrix · Free sandbox signup · TypeScript package

When to use Paybond

Reach for Paybond Kit whenever an AI agent can move money and you need spend that stays inside a budget — not just token metering: agentic payments and commerce, tool-call spend limits, outcome evidence before release, and audit-ready records — across OpenAI, Gemini, Claude/Anthropic, LangGraph, CrewAI, MCP hosts, and custom runtimes, settling through the payment rails your tenant already uses.

• Use Paybond Kit when spend needs authorization against a bounded budget, proof-gated release/refund/dispute handling, or audit-ready receipts.
• Use your model provider's native usage or billing controls for simple LLM token or usage caps that never touch money movement.

Start here (sandbox in ~60 seconds)

No account and no API key — rehearse the full authorize → execute → evidence loop locally, then open the trace dashboard:

pip install paybond-kit
paybond dev loop --offline
paybond dev trace

paybond dev trace renders the middleware timeline at http://127.0.0.1:9477.

To try the same flow against the real hosted sandbox Gateway, first create a free Free Developer workspace if you do not already have a Paybond account (no card required), then:

paybond-kit-login
paybond-kit-init --preset paid-tool-guard --framework provider-agnostic --out paybond_paid_tool_guard.py
paybond agent sandbox smoke \
  --operation travel.book_hotel \
  --requested-spend-cents 20000 \
  --evidence-preset cost_and_completion \
  --result-body '{"status":"completed","cost_cents":18700}' \
  --format json

paybond-kit-login opens a browser device-approval step; the workspace owner (you, right after signup) approves it and the CLI writes a sandbox PAYBOND_API_KEY to .env.local.

Launch the tenant-bound MCP server for MCP hosts (after paybond-kit-login):

pip install "paybond-kit[mcp]"
paybond-mcp-server

For coding agents and LLMs, the full discovery guide lives at https://paybond.ai/llms.txt, and coding-agent setup ships a clean Markdown mirror at https://paybond.ai/docs/kit/coding-agent-setup.md.

Install

Core SDK:

pip install paybond-kit

Optional integrations — install only the extras your runtime needs:

pip install "paybond-kit[langgraph]"
pip install "paybond-kit[claude-agents]"
pip install "paybond-kit[openai-agents]"
pip install "paybond-kit[crewai]"
pip install "paybond-kit[mcp]"
pip install "paybond-kit[langgraph,mcp]"

pipx: quote extras on zsh. Install with pipx install 'paybond-kit[langgraph]'. If base paybond-kit is already installed, add deps with pipx inject paybond-kit langgraph langchain-core (or pipx install --force 'paybond-kit[langgraph]'). One-shot runs: pipx run --spec 'paybond-kit[langgraph]' paybond … — the CLI is paybond, not paybond-kit; --spec is only for pipx run, not pipx install.

Runtime-neutral guard helpers, policy files, and paybond agent sandbox smoke are included in the core package. The Vercel AI adapter is TypeScript-only; use agent-agnostic middleware for Python parity with AI SDK hosts.

Open source and supply chain

paybond-kit is distributed as open-source software under the Apache 2.0 license. The source repo and published artifacts include the full license text in LICENSE. Tagged releases publish through PyPI Trusted Publishing (OIDC, no long-lived tokens) with CycloneDX SBOMs generated in CI. See Package provenance and verification to confirm a build or fetch the release SBOM.

Requirements

• Python 3.11+
• A paybond_sk_sandbox_... or paybond_sk_live_... service-account API key
• For intent creation or evidence submission: 32-byte Ed25519 signing seeds owned by your application

Published wheels bundle the paybond_kit._native extension. maturin develop is only required when building from a local checkout.

Create a sandbox key for local development:

paybond-kit-login

paybond-kit-login writes a sandbox PAYBOND_API_KEY to .env.local with file mode 0600, adds the default .env.local target to .gitignore when needed, and refuses to overwrite an existing key unless --force is passed. Custom env-file paths inside a git repo must already be ignored. Live production keys are created by tenant admins in Console and stored in deployment secret managers.

CLI

The package ships the paybond CLI (paybond, paybond-kit-init, paybond-kit-login, paybond-mcp-server).

Scaffold a starter project from bundled templates:

paybond init --template invoice-agent
pip install -r requirements.txt
paybond agent sandbox smoke --policy-file paybond.policy.yaml \
  --operation saas.provision_seat \
  --requested-spend-cents 2900 \
  --evidence-preset cost_and_completion \
  --result-body '{"status":"completed","cost_cents":2900}' \
  --format json

End-to-end sandbox smoke (bind + execute + evidence) with no app code:

paybond agent sandbox smoke \
  --policy-file paybond.policy.yaml \
  --result-body '{"status":"completed","cost_cents":2900}' \
  --format json

Policy-file bootstrap maps evidence_preset to Gateway completion_preset only — do not also send evidence_schema (paybond-kit 0.11.4+). See Agent policy.

agent sandbox smoke only requires paybond-kit. Framework demo commands load their optional extras on demand.

First guardrail scaffold

Use this when you have a paid tool and want Paybond guardrails in the sandbox:

paybond-kit-init \
  --preset paid-tool-guard \
  --framework provider-agnostic \
  --out paybond_paid_tool_guard.py

The generated integration opens Paybond from the environment, loads .env.local when PAYBOND_API_KEY is not already present, bootstraps a sandbox guardrail intent, wraps your paid-tool handler, and submits sandbox evidence. It does not generate a paid-tool implementation. Free Developer is sandbox-only; live settlement rails start on paid production plans.

Tenant isolation

Every session is bound to the tenant realm echoed by gateway-authenticated service-account introspection.

• Do not pass tenant ids by hand for normal SDK usage.
• Construct one Paybond session per tenant/service account.
• Treat any tenant or intent echo mismatch from Harbor as a severity-zero defect.

Quick start

import asyncio
import os

from paybond_kit import Paybond


def required_env(name: str) -> str:
    value = os.environ.get(name)
    if not value:
        raise RuntimeError(f"missing {name}")
    return value


async def main() -> None:
    paybond = await Paybond.open(
        api_key=required_env("PAYBOND_API_KEY"),
        expected_environment="sandbox",
    )
    try:
        print("tenant realm:", paybond.harbor.tenant_id)
    finally:
        await paybond.aclose()


asyncio.run(main())

Agent spend controls

Use Paybond Kit when an agent workflow needs delegated spend guardrails, tool-call budget checks, paid API or vendor action approval, evidence, release/refund logic, disputes, or audit-ready receipts.

import asyncio
import os

from paybond_kit import Paybond


async def main() -> None:
    paybond = await Paybond.open(
        api_key=os.environ["PAYBOND_API_KEY"],
        expected_environment="sandbox",
    )
    try:
        guardrail = await paybond.guardrails.bootstrap_sandbox(
            operation="travel.book_hotel",
            requested_spend_cents=20_000,
            currency="usd",
        )

        guard = paybond.spend_guard(guardrail.intent_id, guardrail.capability_token)
        guarded_tool = guard.guard_tool(
            operation=guardrail.operation,
            requested_spend_cents=guardrail.requested_spend_cents,
            handler=book_hotel,
        )

        result = await guarded_tool({"hotel_id": "hotel_123", "max_price_cents": 20_000})
        await paybond.guardrails.submit_sandbox_evidence(
            guardrail.intent_id,
            {"result": result, "sandbox": True},
        )
    finally:
        await paybond.aclose()


asyncio.run(main())

The paybond.harbor and paybond.guardrails clients are created by Paybond.open(...) and bound to the tenant resolved from the service-account API key. Production integrations read capability_token from paybond.intents.create(...), or from paybond.intents.fund(...) after an x402_usdc_base payment challenge is satisfied.

What the package includes

Core SDK:

• Paybond.open(...) for API-key-only, tenant-derived hosted sessions
• HarborClient for capability verification, intent creation, x402 funding, evidence submission, and ledger reads
• paybond.signal and paybond.fraud on Paybond sessions opened from one service-account API key
• PaybondIntents helpers for principal-side signing, x402 funding, payee-side signing flows, and settlement confirmation
• PaybondSpendGuard, authorize_spend, and guard_tool for spend-named wrappers around capability verification
• Runtime-neutral and framework aliases: paybond_agent_tool_spend_guard, paybond_runtime_neutral_tool_spend_guard, paybond_langgraph_tool_spend_guard, and paybond_mcp_tool_spend_guard
• paybond_runtime_tool_call_adapter for agent SDKs and custom runtimes that expose a tool-call object plus an application-owned executor

Agent middleware and CLI:

• PaybondAgentRun, tool registry, interceptor, and policy-file binding
• paybond init, paybond agent run bind, paybond agent tool execute, and paybond agent sandbox smoke
• Optional LangGraph, Claude Agents, and MCP integrations via extras (see table above)

Gateway and trust helpers:

• GatewaySignalClient and ServiceAccountSignalSession for tenant-scoped Signal reads and signed portfolio artifacts
• GatewayFraudClient and ServiceAccountFraudSession for tenant-scoped fraud assessments, review queues, review events, metrics, and release-gate config
• list_plaid_banks and fund_ach_with_plaid_bank (paybond_kit.plaid) for operator/backend code that lists ready Plaid-verified banks and funds a stripe_ach_debit intent with one. Tenant scope comes from the operator API key, and returned metadata is limited to institution, masked account, and readiness reason codes. These are deliberately not exported from paybond_kit.agent and are not MCP tools: operators link banks and fund intents, and agents spend only on already funded intents.
• Protocol-v2 helpers for mandate verification, replay-safe recognition proof verification, receipt reads, and A2A discovery
• paybond-kit-login for sandbox device approval and local .env.local API-key setup
• paybond-kit-init for generating a Paybond guardrail integration helper

Agent-facing surfaces are model-provider agnostic. Paybond verifies tool operations and tenant scope, not whether a tool call came from OpenAI, Anthropic, Gemini, a local model, or another runtime.

allowed_tools values are your own tool or operation names, not a Paybond-owned catalog. Harbor enforces string matching against whatever names you chose when creating the intent.

settlement_rail on intent creation is a principal-signed rail request. Stripe destinations and x402 receive addresses stay tenant-owned server-side config and are never supplied by the SDK caller.

The protocol-v2 surface is trust-first: signed mandates, recognition proofs, and receipts work across supported settlement adapters instead of treating any single rail as the product boundary.

Gateway-backed protocol helpers raise ProtocolHttpError with parsed error_code and error_message fields when the gateway returns a JSON error envelope. Recognition-gated flows surface unregistered_key, revoked_key, mandate_agent_key_mismatch, and protocol_binding_mismatch explicitly.

What it does not include

• No console UI workflows, and no bank-linking flow: Plaid Link, public_token exchange, access tokens, and Stripe processor tokens stay server-side in the Gateway and are never accepted or returned by Kit
• No bundled LLM or model runtime — bring your own agent framework and install optional extras when needed
• No model-provider-specific MCP wrapper; the MCP server is host-agnostic and works with any MCP-compatible runtime

Source build

For local development from this directory:

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
maturin develop

Use this path when you are editing the package itself or rebuilding the bundled native extension locally.

Docs

• Agent + LLM discovery guide: https://paybond.ai/llms.txt
• Coding-agent setup (Markdown mirror): https://paybond.ai/docs/kit/coding-agent-setup.md
• Long-form docs: https://paybond.ai/docs/kit
• Agent quickstart: https://paybond.ai/docs/kit/quickstart-agent
• One-command guardrails: https://paybond.ai/docs/kit/one-command-guardrails
• Python quickstart: https://paybond.ai/docs/kit/quickstart-python
• Python SDK reference: https://paybond.ai/docs/kit/sdk-reference-python
• Agent integrations: https://paybond.ai/docs/kit/agent-integrations
• Support matrix (languages, frameworks, rails): https://paybond.ai/docs/kit/support-matrix
• Package provenance and verification: https://paybond.ai/docs/kit/package-provenance
• MCP server guide: https://paybond.ai/docs/kit/mcp-server
• Agent runtime tutorial: https://paybond.ai/docs/kit/agent-runtime-tutorial-python
• Python example projects: https://paybond.ai/docs/kit/examples-python
• LangGraph patterns: https://paybond.ai/docs/kit/quickstart-python#agent-framework-integrations
• Free Developer sandbox signup: https://paybond.ai/signup

Release verification

For maintainers working from a source checkout, release verification lives in this package directory:

python3 scripts/verify_release.py

This builds wheel and sdist artifacts, inspects them for stray local files, validates metadata/extras, and smoke-installs the built wheel in a temporary virtual environment.

Publish to PyPI

For maintainers only:

export MATURIN_PYPI_TOKEN="pypi-..."
./scripts/publish_release.sh

This reruns release verification and then publishes the sdist and wheel with maturin publish --non-interactive.