Skip to content
<- All products
// product

AgentFuse

Catch an agent stuck in a loop and put a cap on its calls.

Brand: Tunedness

Visit sitegithub.com/Tunedness/AgentFuse

AgentFuse is a circuit breaker for an agent’s tool calls. It sits between the agent and the MCP servers it uses as a transparent proxy and notices sessions that repeat the same call endlessly or run past their budget. It is one of Tunedness’s open source projects.

Loop detection

A rule tier catches repeated calls, repeated errors and short cycles; it needs no model and no network. An optional embeddings package adds a semantic tier.

Budgets and rules

Each session can have a budget for duration, call count, estimated tokens and estimated dollars. The token and dollar figures are lower-bound estimates based only on tool input and output. Per-tool rules are matched by glob and take one of allow, warn, deny or require_approval. Calls that need approval go to a human through the CLI or a webhook; a reason (--reason) is mandatory.

Observe first, then enforce

The default mode is warn: AgentFuse observes and reports and blocks nothing. enforce turns blocking on. Everything lives in a single YAML file (fusepolicy.yaml), and its JSON Schema is published. The CLI offers wrap, serve, init, validate, report, approve, deny and models install. Telemetry flows as tunedness.* events over OTLP. AgentFuse can be chained with McpGuard; in the chain McpGuard is the outermost proxy.

Installation and scope

AgentFuse is installed from its GitHub repository: you clone it and build it. There is no package published on npm yet, and the agentfuse package on npm does not belong to this project. Node.js 20.19 or later is required. Left out on purpose: a guarded HTTP gateway, an in-process Python SDK, usage ingest, A2A traffic, framework plugins and a cost dashboard.

License

Apache-2.0. The source is developed on GitHub.

// Products

Key features

  • Transparent MCP proxy: no code changes on the agent side or the tool side
  • Loop detection: repeated calls, repeated errors, short cycles; optional semantic tier
  • Per-session budgets for duration, call count, estimated tokens and estimated cost
  • Per-tool rules: allow, warn, deny, require approval
  • Human approval through the CLI or a webhook
  • One YAML policy file and a published JSON Schema
// faq

Frequently asked questions

What does AgentFuse do?+

It is a circuit breaker for an agent's tool calls. It sits between the agent and its MCP servers as a transparent proxy, notices sessions that fall into a loop or exceed their budget, and warns, blocks or asks for human approval according to your rules.

Which agent frameworks does it work with?+

It is framework-agnostic: LangGraph, CrewAI, the Claude Agent SDK or a hand-written loop. As long as the agent reaches its tools over MCP, neither side needs code changes.

Does it start blocking right away?+

No. The default mode is warn: it observes and reports but does not block. Blocking starts with enforce mode.

How is it installed?+

AgentFuse is installed from its GitHub repository: you clone it, build it and run the command-line tool locally. There is no package published on npm yet. The agentfuse package on npm does not belong to this project.

How exact is the cost budget?+

Call count and duration are exact. Token and dollar budgets are lower-bound estimates based only on tool input and output; the project says so explicitly.

What is the license?+

Apache-2.0. It is open source; the code is on GitHub.

Let’s build your next project together

Tell us your idea and we’ll design the right solution for you.

Get in touch