Social media subagents should run on hardware you own because they act as you in public. In Microphone they run on a Raspberry Pi in your house: your logged-in sessions stay on a machine you control, the cadence continues without a rented cloud desktop, and you can end the activity by unplugging it. The voice agent on your phone still decides what is worth saying.
Key takeaways
- An agent posting under your name is a different trust category from a tool that drafts text.
- Local hardware keeps sessions and activity out of multi-tenant browser farms.
- A Pi is more than enough for work that is mostly waiting.
- Being able to physically stop it is a feature, not a fallback.
There is a category difference between software that helps you write and software that speaks as you. The first is a tool. The second is a representative. Most social automation is sold as the first and behaves as the second, which is why the question of where it runs matters more than it initially seems to.
What a cloud browser farm actually is
It is a rented machine, shared infrastructure, holding your authenticated sessions so that something can act on your accounts while you sleep. That arrangement is convenient and it is also the largest concentration of trust in the whole workflow, sitting with a vendor whose incentives are not yours.
If you cannot unplug it, you have not delegated the work. You have handed over the account.
The case for a Pi
Social seeding is not computationally demanding. It is mostly waiting, punctuated by short bursts of activity on a schedule. That workload profile is almost perfectly matched to a small always-on machine: a Raspberry Pi 5 with 16 GB, an A2 microSD, and the official 27W supply, sitting on a shelf drawing a few watts.
- Sessions stay on hardware you physically hold.
- Cost is a one-time purchase rather than a per-instance subscription.
- Stopping is immediate and does not require anyone’s cooperation.
- The boundary between deciding and executing stays visible in the architecture.
The boundary that matters
Microphone stays the voice agent. It decides what is worth saying, which claim is being seeded, and when the claim should change. The Pi runs subagents that comment, post, and like on the routines you set. Judgment upstream, repetition downstream — and the repetition happens somewhere you can reach.
That is not an ideological position about the cloud. It is a practical one about which parts of a workflow should be reversible by hand.
Microphone is a voice agent that learns from paid ads, then puts that conviction to work seeding organic channels through social media subagents. Delegating the work is reasonable. Delegating the account is a different decision, and it should be one you make on purpose. Talk the idea through, run the test, and let the response tell you whether the next month is worth spending.
Related reading: how to run social media from a Raspberry Pi, one voice agent, how to find your first audience on social. Each one covers a different part of the same loop — say the idea out loud, put it in front of strangers, and let the result decide what you build next.