Platform overview

One model, agents inside, running work on the machines you own.

Sentinel is a model for the work chat assistants hand back: its agents read your files, run your tools, control your devices and keep going after you close the laptop. Built on open-weight models. Your rules are enforced, your data stays on hardware you own, and the pricing is on the page.

Explore use cases
getsentinel.ai/app
Session

User

The weekly export landed in ~/Downloads. Clean it, merge it into the master sheet and tell me what changed.

Sentinel

Merged 1,284 rows into master.xlsx. What changed since last week:

  • 37 new customers, 4 closed accounts
  • Two duplicate invoices removed (INV-2291, INV-2304)
  • Revenue up 6.2 % week on week
Model: sentinel · localRuns on your Mac

How it fits together

Three pieces, one platform

A model with agents inside, rules it cannot move, and a runtime you already own.

01

The model

One call in, finished work out. Inside it, agents plan the job, call tools, check their own work and report back with the result. When a step fails they retry, compact or reroute on their own.

02

The rules

Allowed tools, paths and devices, a daily spend cap and an audit log of every action. The rules live outside the model, so no prompt can talk it past them.

03

The runtime

Runs on your Mac, a server you own or a private container. Sessions, memory and keys stay there. Open it from the web, the Mac app or your phone.

Capabilities

What you get

Self-healing turns

Every failure is classified and cured in order: retry, compact, reroute to another model. One line of notice, then the work continues.

Plans you can resume

Long jobs keep a plan. Close the laptop, lose the network, restart the server: the turn resumes from the last finished step.

Memory that learns

After a mistake it writes a lesson so the same mistake does not happen twice. Documents, screens and short videos become knowledge it cites and playbooks it can follow.

Tools and MCP

Files, shell, browser, screenshots and any MCP server. Servers that crash come back on their own with backoff.

Rules and audit

Per-device permissions, path limits, spend caps and a log of every call, so a team can run it without losing control.

Every device

The same model on your laptop, your always-on mini and your phone, with sessions that run side by side without blocking each other.

Live view

Measured, not promised

Numbers from our own test runs against the production setup, with the test that produced each one.

Turns finishedno freezes
0%

Share of turns in the ten-turn smoothness run that ended with a real answer and no error, with a median first output of 1.3 s (tests/test_smoothness.py, 12 Oct 2026).

Second-step prefill
0svs 44 s cold
Prompt reused from cache8,180 tokens
Prefill with the cache0.5 s
Prefill without it44 s
Identical tool call allowed before the loop guard3 times

Private by construction

We keep this much of your work:

A design choice, not a policy line. Sentinel runs on hardware that belongs to you, so there is nothing on our side to log, leak or train on.

0bytes of your work on our servers
Sessions
never stored
Files and outputs
never uploaded
Model keys
never seen
Your work
never trained on

Built for real workflows

Turn a folder of exports into a dataset

Point it at the mess: exports, PDFs, inboxes. Sentinel writes clean, schema-checked rows to the format your pipeline expects, flags the rows it could not read, and keeps the steps so next week is one command.

getsentinel.ai/generate

Describe the output

Read every invoice PDF in ~/Finance/2026 and give me one row per line item.
2,140 rowsCSV + Parquetschema checked

invoices-2026.parquet

0 / 5 parts · writing…

FAQ

Questions, answered

What is Sentinel for?

Work that needs more than a chat window: reading and writing your files, running commands, deploying, researching with sources kept, controlling your own devices, and jobs that take hours.

What will it not do?

It follows your rules: tools, paths and devices you have not allowed are off, and spend caps are enforced outside the model. You stay responsible for what you ask it to do, and our terms rule out using it to harm people or to break into systems you do not own.

Do you store my work?

No. Sentinel runs on your machine or a container that belongs only to you. Sessions, memory and keys live there. Delete everything with one command.

Which models does it use?

Local open models on your own hardware by default, with optional hosted models on paid plans. Sessions keep the same model settings so the prompt cache holds and turns stay fast.

How is it priced?

Free to start, then $20 or $50 a month for more usage. Every plan and limit is on the pricing page, with no sales call.

See how Sentinel handles your workflow.

Free to start, no card needed. Bring a real task and watch it finish.

See pricing

Private beta

Join the waitlist.

Seats open in order. We write once, when yours is ready.