AgentGuard
AgentGuard adds budget enforcement, credential isolation, token metering, and MCP policy controls to AI-agent workflows. Developer teams use its SDK to limit agent costs and govern access to tools and credentials.
One of 617 tools in AI Agents
Who it's for
- AI-agent developers using OpenAI, Anthropic, Google, or OpenAI-compatible APIs who need per-workflow spending limits.
- Platform and security teams managing agents that need isolated credentials or OAuth scope enforcement.
- Teams using MCP tools that need tool-call policies, violation alerts, and audit logs.
How it fits your workflow
Install the Python package and initialize a project. Wrap a supported model client with AgentGuard's `guard(...)` function, then set controls such as a per-run budget, isolated authentication, fallback model, and limit behavior.
As an agent runs, AgentGuard tracks token use, applies budget controls and model downgrades, and can terminate a run at its cap. For MCP calls, Sentinel evaluates tool calls against configured policies before execution.
Teams can use the metering features to track usage across providers and connect Stripe to invoice clients based on token consumption.
Pricing
Free: Guard Core is free forever and open source. It includes core budget controls, CLI cost reports, all-provider support, and up to 10K tracked calls per month.
Guard Core
$0 forever
- · Open source
- · Per-workflow budget caps
- · Automatic model downgrade
- · Hard kill switch
Guard Pro
$49/month
- · Everything in Guard Core
- · 100K tracked calls per month
- · Web dashboard
- · Slack and Discord alerts
Shield
$149/month
- · Everything in Guard Pro
- · Per-agent credential isolation
- · OAuth scope enforcement
- · Agent identity registry
Sentinel
$349/month
- · Everything in Shield
- · MCP tool-call interception
- · Natural-language policies
- · Per-tool access control
Fortress (Enterprise)
Contact for pricing
- · The page lists unlimited calls, SSO, SLA, and GRC integration as Fortress options.
Prices checked on Oct 9, 2026 from the vendor's site. They can change; confirm before you buy.
Key features
- Per-workflow budgets
- Set spending limits per run, user, or feature, with enforcement while the workflow is running.
- Automatic model downgrade
- At 80% of a budget, route work to a cheaper model; the page gives GPT-5.4 to nano and Opus to Haiku as examples.
- Hard kill switch
- At the budget cap, AgentGuard can gracefully terminate a workflow and return a summary.
- Token metering and cost analytics
- Track token consumption and costs across supported providers in real time; the page also describes cost analytics and cost history by plan.
- Usage-based Stripe billing
- Connect Stripe to invoice clients based on actual token usage.
- Rate limits and quotas
- Configure rate limits and spending caps per agent, user, or endpoint.
Works with
- OpenAI
- Anthropic
- OpenAI-compatible APIs
- Ollama
- Stripe
- MCP
Limitations to know
- Tracked-call allowances vary by plan: 10K per month for Guard Core, 100K for Guard Pro, 500K for Shield, and 1M for Sentinel.
- MCP tool-call interception and policy enforcement are listed as Sentinel-tier features.
Alternatives to consider
-
Adversa AI
Choose it if your main need is runtime controls and adversarial testing specifically for coding agents, rather than AgentGuard's combined cost, credential, and MCP controls.
-
Lakera
Choose it if you need an enterprise GenAI security platform covering employee AI usage and risk-based red teaming in addition to protection for autonomous agents.
Getting started
pip install agentguard- Install AgentGuard with `pip install agentguard`.
- Run `agentguard init` to initialize the project.
- Wrap a supported model client with `guard(...)` and configure the controls needed for the workflow.
- Use the initialized SDK with a supported provider client.