Is this model-agnostic?

Yes. Your model provider client stays on the provider API; Paybond wraps tool handlers at the spend boundary.

Do I call submitEvidence manually?

Not when using middleware for registered side-effecting tools. Auto-evidence is required middleware behavior. See the docs for legacy per-tool guard paths.

Sandbox vs production?

Sandbox rehearses authorization and evidence with funded sandbox intents. Production attaches to live intents on your tenant.

How is this different from LiteLLM budgets?

LiteLLM caps model token spend. Paybond authorizes delegated paid work, submits signed evidence, and supports settlement lifecycle (release, refund, dispute).

Agent middleware

Model-agnostic spend middleware for agent tool calls.

Authorize at the tool boundary, submit evidence automatically, and keep one intent per agent run — across OpenAI, Claude, Gemini, LangGraph, MCP, and custom orchestrators.

Start Free Developer Read docs Run sandbox smoke

Why tool boundary, not model boundary

Paybond does not proxy LLM inference. Model providers handle token usage; Paybond authorizes paid and side-effecting tools before they execute and records proof afterward.

Dimension	Paybond middleware	Without Paybond middleware
Spend scope	Paid tools, vendor APIs, bookings, settlement workflows	Model token budgets (LiteLLM, provider dashboards)
Integration point	Tool handler / MCP tool / graph node	Model API proxy or gateway
Evidence	Automatic on successful side-effecting calls	Usage logs or invoices
Authorization model	Capability-scoped authorization — intent bound per agent run	Pre-call if-check in app (easy to bypass per tool)

Sandbox smoke (CLI)

Rehearse spend authorization and evidence in sandbox without writing integration code. JSON output for coding agents.

paybond login
paybond agent sandbox smoke \
  --operation paid-tool \
  --requested-spend-cents 100 \
  --evidence-preset cost_and_completion \
  --result-body '{"status":"ok","cost_cents":100}' \
  --format json

Paybond CLI →SDK reference →

Clone a starter template

Public GitHub template repos ship paybond.policy.yaml, .env.example, npm run smoke (policy-file based), and CI. git clone → paybond login → npm run smoke in under a minute.

Scaffold locally

paybond init --template travel-agent --framework langgraph

Or clone from GitHub

git clone https://github.com/nonameuserd/paybond-travel-agent.git
cd paybond-travel-agent
cp .env.example .env.local
paybond login
npm install
npm run smoke

Travel booking agent
langgraph · travel
npm install && npm run smoke
Use template on GitHub →
Shopping checkout agent
vercel-ai · shopping
npm install && npm run smoke
Use template on GitHub →
SaaS provisioning agent
openai-agents · saas
npm install && npm run smoke
Use template on GitHub →
OpenAI shopping agent
openai-agents · shopping
npm install && npm run smoke
Use template on GitHub →
Claude travel agent
claude-agents · travel
npm install && npm run smoke
Use template on GitHub →
MCP coding agent
mcp · custom
npm install && npm run smoke
Use template on GitHub →
Invoice agent (Python)
langgraph · saas
pip install -r requirements.txt && npm run smoke
Use template on GitHub →

Framework adapters

Bilingual frameworks ship TypeScript and Python docs, exports, and smoke on one page — complete sandbox smoke in either language without reading the other kit.

Agent-agnostic core
Defaultavailable
TypeScriptPython
Default recommended when you have no framework SDK — custom orchestrators, Claude/Gemini tool-use, or queue workers. createPaybondGenericAgentConfig in TypeScript and Python with identical semantics.
```
paybond agent demo generic smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
LangGraph
available
TypeScriptPython
ToolNode hooks in both kits — paybondAwrapToolCall / paybond_awrap_tool_call and paybondToolNode parity. Bilingual docs and smoke; no need to read the other kit.
```
paybond agent demo langgraph smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
Claude Agents SDK
available
TypeScriptPython
In-process MCP custom tools wrapped with Paybond middleware. TypeScript (@anthropic-ai/claude-agent-sdk) and Python (claude-agent-sdk) parity — bilingual docs and smoke.
```
paybond agent demo claude-agents smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
Vercel AI SDK
available
TypeScript
toolApproval + paybondVercelWrapTools for AI SDK generateText / streamText. TypeScript only — Python Kit links here.
```
paybond agent demo vercel-ai smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
OpenAI Agents SDK
available
TypeScript
Input guardrails + wrapped invoke for OpenAI Agents function tools. TypeScript only — Python Kit links here.
```
paybond agent demo openai-agents smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
MCP hosts
available
TypeScriptPython
Tenant-bound stdio paybond-mcp-server for Claude Desktop, Codex, and generic MCP orchestrators. TypeScript and Python kits ship the same tool surface.
```
paybond login
# Launch: npx -y -p @paybond/kit paybond-mcp-server
# Or see /docs/kit/coding-agent-setup
```
Adapter docs →
Mastra
planned
TypeScript
Native Mastra runner helper is planned. Use agent-agnostic wrapTools or MCP today — request the adapter if Mastra is your primary stack.
```
paybond agent demo generic smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
CrewAI
planned
Python
Native CrewAI runner helper is planned. Use agent-agnostic wrap_tools or MCP today — request the adapter if CrewAI is your primary stack.
```
paybond agent demo generic smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →
Cloudflare Agents
planned
TypeScript
Native Cloudflare Agents runner helper is planned. Use agent-agnostic wrapTools or MCP today — request the adapter if Cloudflare Agents is your primary stack.
```
paybond agent demo generic smoke --operation paid-tool --requested-spend-cents 100 --evidence-preset cost_and_completion --format json
```
Adapter docs →

Full install surfaces and rails: Kit support matrix

FAQ

Is this model-agnostic?: Yes. Your model provider client stays on the provider API; Paybond wraps tool handlers at the spend boundary.
Do I call submitEvidence manually?: Not when using middleware for registered side-effecting tools. Auto-evidence is required middleware behavior. See the docs for legacy per-tool guard paths.
Sandbox vs production?: Sandbox rehearses authorization and evidence with funded sandbox intents. Production attaches to live intents on your tenant.
How is this different from LiteLLM budgets?: LiteLLM caps model token spend. Paybond authorizes delegated paid work, submits signed evidence, and supports settlement lifecycle (release, refund, dispute).