Share to action · local by default

Your shared screenshot
becomes useful.

Share a screenshot, receipt, message, or document. Marmot understands it locally, proposes an editable action card, and waits for your approval before adding a calendar event, reminder, draft reply, or saved note. Nothing leaves the device.

Android APK · iOS builds from source · MIT · works in airplane mode
Local chat result Share-to-action screen with local action choices Marmot logo
No account. No cloud. No telemetry.
$ network_requests --after-download → 0  // verified by architecture, not policy
One useful loop

Share it. Understand it.
Approve the outcome.

Share a screenshot or message → understand it locally → propose the next action → approve and execute it on the phone.

Small local models handle the interpretation; deterministic phone code handles permissions, dates, and writes.

01 · Share

Bring the moment in.

Send Marmot a screenshot, receipt, message, article, or document from the app you are already using.

02 · Understand locally

See the useful facts.

Extraction happens on the phone, with a clear local-only proof and a typed preview you can edit before anything changes.

03 · Approve → act

Finish in one tap.

Approve a calendar event, reminder, draft reply, or saved note. Marmot never sends or writes behind your back.

Captured, not promised

Two verified boundaries
on the phone.

These are Android runtime captures from the canonical Pixel 7 / Android 35 baseline. Screenshot intake and private local vision processing are verified; the separate text-share path reaches an approved calendar write and undo. Screenshot OCR quality remains an honest open gate.

External share

The moment arrives privately.

Marmot receives the shared image and copies it into its private attachment storage before local processing.

Marmot Quick actions showing an image received from an external Android share
Text share · approve, then act

The phone waits for you.

For shared text, the calendar preview shows the resolved time. The write happens only after approval and exposes an undo path.

Marmot calendar action card showing an approved event and Undo event

Read the runtime log and limitations → · See sanitized before/after examples

More in Labs

Power for later,
when you need it.

Agent mode, voice, memory, research, MCP, and repo tools remain available for power users. They are Labs surfaces; the share-to-action loop is the product's first promise.

Agent mode

It plans. It uses tools. It checks its work.

A planner decomposes the task, fresh-context subagents execute each step with calculator, clock, chat search, and RAG tools, live ☑ check-offs tick by, and an optional judge scores the answer before you see it.

Agent mode screenshot
Voice

Talk to it. Bring it to meetings.

A hands-free conversation loop with an animated voice stage — plus meeting mode: continuous transcription, and say “Marmot, …” for a suggested contribution it will only speak when you tap.

Voice mode screenshot
Memory + RAG

It remembers — by meaning.

Facts about you, episodic summaries, imported documents, even whole GitHub repos — recalled with on-device embeddings and injected into every agent run.

Web research · opt-in

Research with receipts.

Flip one switch and the agent can search and read the web, citing sources. Leave it off and Marmot is provably offline.

Your data

Yours to carry away.

Markdown sharing, JSON backups to Drive or OneDrive via the share sheet, and safe re-import that never overwrites newer history. No cloud SDKs. No OAuth.

Labs · under the hood

The whole frontier-agent stack, in your pocket.

Nine layers, all shipped, all unit-tested — the same architecture the big coding agents converged on, running against a 2–4B model on your phone.

system SYSTEM.md kernel + user personas
skills trigger → procedure registry
loop observe → decide → act → verify
tools calc · clock · chats · docs · web
memory user·project·episodic·semantic
planning live plan panel, model-driven ☑
reflection self-critique may revise
judge independent scoring, fail-closed
subagents fresh-context executor per step
192
unit tests, CI-enforced
8.6
text tok/s baseline
~40s
vision baseline on emulator
The catalog

One champion per weight class.

The highest-rated open model at every size a phone can carry — byte-exact, benchmark-picked, re-evaluated as new models ship. Or import any .gguf of your own.

🪶

Qwen3.5 0.8B

0.53 GB

Instant on any phone. 200+ languages.

🥊

Qwen3.5 2B

1.28 GB

The top-rated small all-rounder.

🥋

SmolLM3 3B

1.92 GB

Strongest fully-open 3B.

🏋️

Qwen3.5 4B

2.74 GB

Strongest dense 4B. Wants 8 GB RAM.

👑

Gemma 4 E4B

4.06 GB

Closest to cloud quality on a phone.

all Apache 2.0 · resumable background downloads · RAM-fit badges from your device’s actual memory
~2 MB
JS bundle
1.2–1.7s
voice first-audio target
0.34%
logo pixel diff, traced
MIT
forever