Browse the documentation

Docs Privacy, limits and help

Alpha limits

Geometer 0.1.0-alpha.2 is unfinished. This page lists what it does not do, what has been tested and what has not, so you can decide how far to rely on it.

Stuck? Book 20 minutes A person replies within two working hours.

On this page

Platforms#

  • Apple silicon Macs only on macOS. No Intel Macs, no Windows, and no Linux desktop app. On Linux, geo open opens the workspace in your browser.
  • Not signed by Apple. macOS asks you to allow the app once for each version. A signed and notarized app is planned for the main release.
  • macOS versions. Tested on macOS 26. It is built for macOS 12 and later, which has not been tested.
  • No automatic updates. There is no update check. To upgrade, run the newer installer. See Install.

What has and has not been tested#

Geometer has been used by its own team only. No outside person has installed it without help. Your reports decide what is fixed first.

  • On a Mac, earlier release candidates of the terminal build were installed, upgraded, rolled back and uninstalled, with real checks of the sample projects. Those runs were made before the changes since, so they show what held on that date.
  • The Linux builds (x86_64 and arm64) were installed, upgraded, rolled back and uninstalled in containers, and geo status and geo ci ran there. Those containers had no Docker, so no check was run in them. Live checks on Linux have so far been run from source builds and not from the released files. The x86_64 build was tested under emulation.
  • The workspace was driven in a real Chromium browser on a Mac, in runs made before the latest changes. Safari and Firefox were not tested. The Mac app's window itself was checked through its server and its sign-in; clicks and the folder picker in that window were not driven by automation.
  • Starting a stopped Docker Desktop or Colima from the Docker card or with geo doctor --fix, installing Colima with Homebrew, and downloading Ollama and its model on a computer without them have not been tested for real.
  • The framework support listed on Works with was tested on small agents written for the purpose, one for each framework, and on Geometer's own open agents. Most public example projects of those frameworks that were tried could not run for their own reasons, such as an old dependency that does not install, and the check said so instead of passing them.
  • geo mcp wrap, geo mcp try, geo mcp record and geo mcp twin were tried against the simulated services and small agents, never against the real services. Recording a model's answers through an https provider was tested only against a local stand-in, not against a real provider.
  • The setting for your coding assistant was tried with Claude Code. The Cursor setting follows Cursor's documentation and has not been tried in Cursor itself.
  • The geo command carried inside the Mac app, and the Add the geo command link, have not yet been checked from a released app on a computer other than Geometer's own.
  • The cases were written and run by the people who built Geometer. They show that the checked behaviour worked on that machine on that day.

Which agents it can check#

  • Geometer checks agents whose tools it can find and simulate. When it cannot, it says what is missing instead of reporting a clean result.
  • Outside services are simulated. A check shows how your agent behaves against those stand-ins, not against the real services.
  • Your agent's model is answered by a stand-in, so a check shows how your application behaves, not how a real model behaves. Check a model change and Check constraint stability can use a real model you choose.
  • For agents whose tools run inside the agent's own process, Geometer observes and controls the tool calls through the framework's hooks. Their effects stay unconfirmed unless they reach a service that Geometer's simulated tools answer.
  • Works with lists each framework as Supported, Partial or Not yet. Supported means a real agent of that kind was checked from start to finish, not that every agent of that kind works. Java, .NET and Go frameworks are Not yet, and MCP servers started with npx or uvx are Partial.
  • Hosted tools (a provider's web search or code execution) and the built-in tools of the Claude Code process behind the Claude Agent SDK are not observed. A turn's parallel tool calls are let through one at a time, so behaviour that depends on tools running together is not reproduced.
  • A framework version the hooks were not tested with still runs, and the report says it was not tested.

Faults and situations#

  • The fault kinds are a lost reply, a lost request, a partial effect, a lost acknowledgement, and a crash and restart. Duplicate delivery of a request is not modelled, and the grid of faults tried's Beyond the map list names the other things it does not model.
  • A crash is tried at most once for each run and only where the agent is waiting at a call, and the agent restarts at once with the same command.
  • A partial effect applies only to a call that declares its parts, and is simulated by the recording tool double.
  • A lost request is seen by the agent as a closed connection.
  • The grid of faults tried and the conditions are summaries of what was stored. Each says what it did not look at. Coverage is shown as cells and lists, never as a score.

MCP, simulated services and twins#

  • The ready-made simulated services are the payments pack (Stripe and Adyen), the commerce pack (Shopify, Zendesk and Intercom) and the CRM and ERP pack (Salesforce, HubSpot and NetSuite). Each is a simulation modelled on the provider's public documentation, never the real service, and it covers only part of what the provider offers. The report says "simulated" wherever the provider appears.
  • A try uses one pack at a time, and a call to a tool the pack does not simulate gets an error that says so. The partial-effect fault is not available in geo mcp try, because no tool of the payments pack takes a list of parts.
  • A try is a live look. Your agent and your model are not controlled, so a second try can go differently. A kept try replays the recorded model answers exactly, so an agent whose prompt holds a value that changes from run to run (a time, a random id) will not match its recording, and the replay says so.
  • A mode is read when a wrapped server starts, so restart your agent or assistant after wrapping, trying or recording.
  • A twin knows only what the recorded session showed. A call it never saw gets "not simulated: not in the recording", and a check with such a call says it could not tell. In geo check a twin answers an MCP server reached by an http address, not a stdio server. Recording hides secrets by their shape (keys, tokens, card numbers, email addresses), so read a recording before you share it. While recording, the real server is called, with real effects.
  • Servers defined in code (MCPServerStdio, MCPClient, MultiServerMCPClient) are not edited by geo mcp wrap, which prints the change to make by hand. See Put Geometer in front of your MCP servers.

Guard#

  • Guard is for trying in development and staging. It has not been qualified for production traffic.
  • Guard only sees calls that go through it.
  • After geo guard stop, Guard takes a few seconds to finish the calls it holds. A geo guard start in those seconds can report a Guard as already running and then leave it stopped. Wait a few seconds between stop and start, and check with geo guard status.
  • It runs as its own program. It starts again after a crash only if you started it with --supervise or as a login item.
  • Evidence of Guard from one computer does not show that it works with a provider or at production scale.

Authority and certificates#

  • Names you type, such as who granted a level, are stored as typed. This computer has no accounts, so Geometer cannot verify who you are.
  • A level you grant takes effect in Guard when Guard restarts.
  • Certificates are signed with a key made on your computer, so they show integrity, not who issued them. Nobody else has checked them.
  • Geometer's hosted models are not available yet. Certificate links are built but not switched on (no service is set up yet). A hosted trust page, shared sign-offs, reviewer comments and questionnaire answers are not part of the alpha.
  • The browser verifier on this site checks the integrity of evidence packs. For certificates, use the open geo-verify. See Verify records yourself.

Terms#

Geometer's binaries are free to use under the alpha terms. The software is unfinished: it may change, stop working, or lose or corrupt its own records between versions. Keep anything important elsewhere. A finding, a clean result or a certificate from Geometer is evidence about what was checked. It does not promise anything about your agent's behaviour.

Release notes and downloads#

Something unclear, wrong or missing on this page? Report a problem with these docs.