Spend governance
that never sleeps.

AgentGuard Spend evaluates policy in your process before a wrapped provider call is dispatched. In enforce mode it can block the call before the provider charges you. With signing keys configured, decisions can be recorded in an Ed25519-signed, hash-chained log.

Node.js

npm install @agentguard-run/spend

Python

pip install agentguard-spend

Start here

Codex and ChatGPT Work plugin →

Install, review hook trust, configure matter policies, and verify signed tool records.

Hosted MCP tools →

Connect Claude, ChatGPT or Cursor to free receipt verification, trial requests and documentation answers.

Install, wrap your provider client, and set a daily cap.

Capabilities →

Every cap window, action, scope dimension, capability tier, and locale.

One budget envelope, resumable checkpoints, and a signed chain for multi-step runs.

Integrations →

Native provider bindings plus the framework adapters exported by the TypeScript package.

Outcome Builder Guide →

Prompt-first setup for law, accounting, insurance, real estate, and e-commerce teams.

What AgentGuard does

Per-call, per-minute, per-hour, per-day, and per-month windows. Store-backed reservations can be atomic.

Block / downgrade / shadow →

Hard block, route to a cheaper model, or just log. Per-cap configurable.

Configured scope →

Build one policy scope key from tenant, team, user, agent, task, and provider dimensions.

Cryptographic receipts →

Configure Ed25519 keys to append SHA-256 hash-chained records that verify offline.

Tamper-evident log →

Use the exported in-memory or NDJSON decision-log stores, or implement the store interface.

en-US · es-419 · pt-BR →

Native-locale error messages and blocked traces. Auto-detected.

Integrations

LangChain.js →

An awaited callback handler runs policy preflight before model dispatch.

Wrap an OpenAI-compatible client or guarded fetch call.

Vercel AI SDK →

Middleware preflights generate and stream calls, then settles reported usage.

Tool-event hooks enforce capability policy and record redacted metadata.

Claude Code →

PreToolUse and PostToolUse hooks gate tools and record content-free metadata.

Hermes Kanban →

Task envelopes, capability ceilings, per-task caps, and parent-linked DAG receipts.

Provider bindings →

OpenAI-compatible chat completions, Anthropic Messages, and Bedrock InvokeModel.

Reference

npm package ↗

@agentguard-run/spend · v0.19.0 · package license permits production up to 10,000 enforcement calls per calendar month. Patent-pending.

MCP package ↗

@agentguard-run/mcp · v0.3.0 · signed spend decisions for MCP hosts.

Burn package ↗

@agentguard-run/burn · v0.2.6 · local session limits for agent runtimes.

PyPI package ↗

agentguard-spend · Python package reference.

Source, issues, examples. Star to follow releases.

Full SDK reference in one file. Paste into Claude / Cursor / GPT.

Need something custom?

Need an integration that is not in the exported TypeScript surface? Request direct engineering contact.