Skip to content
Northbeams Labs

Tool

agentsurface

A local inventory of the agent machinery installed on a machine. There is one question behind it: what agent software is on this computer, and who published it.

Status
Version 0.1.0, released 31 July 2026
Published by
Northbeams Labs
Runs
Locally, on macOS, Linux and Windows

What it inventories

  • MCP servers, across every client it can find rather than a few fixed paths. User scope and project scope, on macOS, Linux and Windows.
  • Desktop extensions, plugins, marketplace plugins, skills, connectors and scheduled agent tasks.
  • Instruction files, the full set rather than one vendor’s.
  • AI-aware browser extensions. It enumerates browser profiles locally, then matches what it finds against the catalogue Northbeams maintains.
  • Tool-description drift. It keeps a hashed local baseline of tool names and descriptions, so a later run can flag a definition that changed under you.
  • Named findings, in the sense that each item says what it is, who published it, and what it can reach.

What this does not do

These are refusals rather than defaults. None of them is a setting, and there is no flag that turns any of them on.

  • No network calls of any kind. Not for updates, not for catalogue lookups, not for anything.
  • No account, no sign-in, no token.
  • No upload. Findings stay on the machine that produced them, and nothing is sent to Northbeams or to anyone else.
  • No telemetry, and no lead capture: at no point does it ask for an email address.
  • No model calls. Nothing is sent to a language model, tool names and descriptions included.

This is enforced rather than promised: a build check fails the release if the binary imports the network library.

What it does not detect

Written before the code, so the limits are not something you discover afterwards.

  • It inventories what is installed. Whether any of it is malicious is a judgement it does not make.
  • It does not detect prompt injection, and does not pretend to.
  • It reads local files only, so anything running purely in a cloud account is invisible to it.
  • Catalogue matching is only as fresh as the snapshot compiled into the binary you are running.

What it prints

  • A readable summary in the terminal, by default.
  • --json for machine-readable output.
  • An explicit list of what the run did not look at. That prints on every run.
  • No risk score. A score we cannot defend line by line is a claim, and claims without data are the thing this imprint exists not to make.

Installing

Both commands work. Version 0.1.0 is released for macOS and Linux. There is no Windows build yet.

brew install northbeams-labs/tap/agentsurface
go install github.com/Northbeams-Labs/agentsurface/cmd/agentsurface@latest

Neither command needs an account, and neither one registers anything with us. Every release carries a signed checksums file, a build provenance attestation and a bill of materials. The steps to check them yourself are in docs/VERIFY.md in the repository.

Not in the first version

Absent on purpose. All three are reasonable to add later:

  • SARIF output.
  • CVE lookup.
  • A signed local receipt of a run.