Pre-use questions

[smoochy]

Hey.
Thanks for bringing this project to life.

I'm very keen on finding out what of my tasks and conversations have the most usage to improve on my prompts or working habits.

For private, I would use this in combination with Codex.

I am also thinking of using the tool in the work environment. There we use Claude with Teams Plan.

Therefore I do have 2 questions

• most importantly: can you make a statement about data privacy? Is there any data leaving the proxy server? Is there any data transmitted to any other device / party other than the local machine?
• what is the impact on tokens, as a MCP is being used? Did you do some benchmarks comparing calls and their costs with and without your tool being activated? I'm just asking as the 5h and 7d is a big topic in our company.

Thank you very much and keep up the work.

Kind Regards,
smoo

[marmutapp]

Hey @smoo, thank you for the kind words and for raising these questions. They are important, especially for team / work-environment usage.

1. Data privacy

By default, Observer runs locally. Nothing leaves your machine unless you explicitly enable routing AI-provider traffic through Observer’s optional proxy, which is used for conversation compression. Even In the case of Observer proxy, the traffic is only sent to the AI provider you were already calling.

We currently do not use telemetry, analytics, update checks, CDN-hosted assets, remote fonts, Sentry, OpenTelemetry, Google Analytics/gtag, cloud sync, remote backups, or shared storage. The local dashboard binds to 127.0.0.1:8080, so it is loopback-only.

The narrow exceptions are all opt-in and explicit.

We're also preparing a formal data privacy statement and will share it within the week. For the planned team-views/admin panel, the design will be based on opt-in OpenTelemetry export, with a documented data-flow diagram before anything ships. Nothing will leave a workstation unless an admin explicitly enables it and points it to a team collector.

2. MCP, compression, and token impact

We’re currently working on a benchmark comparing real-world token usage with and without the MCP/proxy setup enabled (MCP is disabled by default). We plan to publish the results within this week as well. We just worked on putting together a comparison for Codex models across effort levels and are doing a similar one for Claude models. Compression efficacy is also scheduled for this week itself.

From our own usage so far, we’ve seen savings in the range of roughly 5-25% when compression is enabled, but that is based on a limited sample. We want to expand the benchmark before making stronger claims.

You can expect a measured A/B comparison in an upcoming release.

Thanks again for bringing this up. Both privacy and token impact answers will be folded into the docs as part of the 1.7 line.

[smoochy]

Thank you very much for your detailed explanation.

Regarding token usage: I wasn't thinking of your tool decreasing token usage (with its compression) but actually if you noticed increased token usage by using the tool - no matter if with compression or without.

In the past I made the experience that using additional stuff (with or without an MCP) increased token usage. And that's what I would like to prevent. As said we crave with our 5h and 7d limits 😀

Cheers,
smoo

[marmutapp]

Oh. I think I misunderstood.

Let me clarify.

Observer doesn't increase token usage at all. It doesn't operate as an MCP - Observer only logs your usage (token, tool & event).

Only if you enable the MCP, for which there is an overhead of tool definitions of ~1800 tokens (even the MCP will reduce total token usage for the session but this is a completely opt-in feature for the user), will there be an active change in your token usage.

[marmutapp]

As an update, we have done the following since the last iteration:

1. introduced privacy statement where we clearly say that we do not do any telemetry whatsoever.
2. We have updated some compression stats - we're seeing 5-25% compression across both GPT 5.4 and Opus 4.7 models - we didn't use GPT 5.5 for testing yet. But we are planning that as well this week.
3. introduced a beta teams - still a limited beta. It is available but we are still running tests.

Sharing for full transparency.