Reviewing SwarmMarshal? Start here.

The claim: answers aren't plausible-sounding text — every important claim links to the exact email or message it came from, and the app says "not in the record" rather than guessing.

Verify it: ask something only your mail knows ("what did the plumber quote me?"), then click the citation and read the original message. Then ask something your mail can't know and watch it decline instead of inventing.

The claim: the context layer belongs to the user, so the model is a plug-in part. The same question works on GPT, Claude, Gemini, or a free local model, with the same citations.

Verify it: ask a question on a cloud model, flip the assistant to a local Ollama model, ask again. Same grounded answer, zero cloud traffic — you can pull the network cable first if you want the theatrical version.

The claim: iMessage has no API — a cloud assistant simply can't see it — and corporate IT rarely lets a SaaS AI ingest the company mailbox. A local app can do both, with the user's own credentials.

Verify it: on a Mac, connect Apple Messages and ask a question that spans a text conversation and an email thread. Watch one answer cite both channels.

The claim: for repeated jobs, the AI authors a small, typed, sandboxed C# tool once — then the job runs as deterministic code. A portfolio tracker's refresh dropped from ~42 model calls to zero.

Verify it: in Vibes, describe a stock or weather tracker, let it build, then open the app's “Inside” view: read the generated tool's code, its declared parameters, and its last-run record — and refresh the rows without a single model call.

The claim: work pinned local-only can never fall back to a cloud model (it stops and says so instead); unproven models can't silently take over the mail pipeline (calibration-gated routing); paying per message is off by default; and the docs state plainly that Claude Code subscription routes do not count as "local."

Verify it: read the white paper and routing page, then check the behavior in the app — set a function to local-only with no local model installed and watch it stop instead of silently going to the cloud.

The claim: if you've covered the self-hosted agent-gateway wave, you know the failure modes — exposed ports, token-burning heartbeats, surprise bills. SwarmMarshal's workers are the containment-first version: each is a real hire running in its own supervised OS process, bound to the machine that hired it, talking only to paired same-owner devices over signed local traffic. There is no public gateway and no always-on LLM loop — and paired machines share compute, so a PC can run a local model on the Mac's GPU.

Verify it: hire a worker in chat and watch it wait for your explicit approval before activating. Open your OS process list and see the worker as its own process. Then check your firewall: nothing is listening for the internet. The workers page maps the whole system.

On the record: “I started building SwarmMarshal because I found the agent gateways everyone was excited about unusable. This is the version I could trust on my own machines.” — Scott Crossen, creator of SwarmMarshal. Quote freely.