A new category of software

Stop rentingsoftware.Prompt your own.

PocketContext is a stack for made-to-measure applications. One prompt becomes one executable with the client, server, database, tunnel and HTTPS already decided. You bring the idea.

Illustration of an agent turning a prompt into an application: Make a new product from pocketcontext, with this data model: deals, contacts and follow-ups. Release verified against checksums. Every identity item renamed. Data model replaced end to end. All product checks passing. dealbook, one file · ready to run.
  • Seat pricing
  • Nginx
  • Certbot
  • Postgres to babysit
  • Redis
  • Kubernetes
  • A vendor’s roadmap
  • Docker sprawl
  • A weekend of YAML

The problem

You’ve tried the three
obvious answers.

01

Rent another SaaS.

Priced per seat, designed for the average customer, with your data living on someone else’s roadmap.

You adapt to the software.
02

Lean on one bot.

It can write the plan. But a bot on its own has nowhere to keep the work: no shared records, no board, no review, no history.

Great talker. No workplace.
03

Assemble an agent framework.

OpenClaw, Hermes and friends are powerful, and yours to run: a runtime to update, secure and keep alive, with channels, providers and skills to configure.

You become its sysadmin.

The fourth answer

Build the small app that does exactly your job. Then put an AI team to work inside it.

Anyone can have software made to measure, as long as they don’t have to make every infrastructure decision again.

The trick

A custom app should be only your business logic.

The questions that eat the first month have been answered once, tested, and moved into a shared library. Your prompt starts after them.

  • Which database?SQLite, embedded
  • How do people sign in?Password, email code, OAuth, social. Your call
  • Who renews the TLS certificate?Let’s Encrypt, automatically
  • Which reverse proxy?None needed
  • How do I reach it remotely?Private tunnel, built in
  • How do agents connect?Skill, CLI or MCP with OAuth
  • What is exposed publicly?An allowlist. The rest is denied
  • Is it safe in a synced folder?Atomic, complete copies only

One executable

It does everything.
It’s one file.

  1. Web appDESKTOP · MOBILE BROWSER

    Use and manage your app from any browser. It ships inside the executable, with nothing else to deploy.

  2. Agent skillCLI · CLAUDE · CODEX · ANTIGRAVITY

    Installs itself with the command line your agent drives, on this computer or a paired one.

  3. MCP + OAuthREMOTE AGENTS

    Connect cloud assistants with browser consent, not copied secrets.

  4. HTTPSLET’S ENCRYPT

    Native TLS with automatic certificates. No proxy in front.

  5. TunnelTAILCAT

    Reach your server from your other computers. No domain, no open port.

  6. ServerPOCKETBASE

    The application server, with accounts, API and migrations.

  7. DatabaseSQLITE

    Embedded. No separate service to install, patch or wake up for.

Built for agents

Any data model. Three ways in.
A fraction of the tokens.

REST API

Write with rules

Every collection gets an API, so changes go through validation instead of straight into tables.

POST /api/collections/deals/records
Read-only SQL

Ask, don’t page through

Joins, counts and aggregates in one statement. The agent gets the answer, not the records it would have to read to work it out.

SELECT stage, count(*) FROM deals GROUP BY stage
BM25 search

Find, then read

Ranked full-text search returns IDs and short excerpts, so the agent opens only what matters.

search "acme renewal"

Y Combinator showed that read-only SQL is the interface agents need. PocketContext solved row-level security on top of it.

5MCP tools make up Crew’s whole surface. Every tool an agent loads costs context before it does any work.
273×fewer bytes to find ready work for an agent
611×fewer bytes for counts by status and assignee

Crew 0.7.0 synthetic retrieval benchmark, 500 notes and 300 tasks, identical answers compared with reading the records. These numbers describe those runs.

Access, to measure

Just you, or a whole company.
Pay only for the complexity you need.

Just you

Skip the teams, seats and roles.

Build it for one person, and that’s all it is. You read and write everything.

Agents

Allowlisted tables and columns.

SQLite’s authorizer blocks every other table and column inside the engine, on every query.

Many users

Sign-in your way, row by row.

Password, email code, OAuth or social login, with different data for each role.

A database for every reader.

Writes go to one database. For each user, your filters build a private read database that holds only the rows that user may see, with its own search index.

A query can’t leak a row that isn’t in the file. New copies build in the background and swap in atomically, so reads keep running while the next one builds.

It works because the records and every index live in one system. Spread across services, this takes a team writing policies. Here, your agent writes the filters with your data model.

Writesapp.db
alice.dbteam rows only
bob.dbown rows only
hr.dball rows

Filtered snapshots are part of Core, and PeopleContext, an HR app with manager hierarchies, shows them at work. Copies stay current by default. With background refresh, a permission change reaches reads when the next copy is ready. Read the architecture

Same binary, two lives

A private tool or a real service.
You don’t rebuild to switch.

Your other computerPhone, browser, cloud assistantSame binary, as clientNothing to install
tailcat tunnelHTTPS · MCP + OAuth
Your computeryour-domain.comThe binary + its data

Listens on localhost only. Nothing to buy, no port to open. Your agent works through the CLI and the installed skill.

Native HTTPS on ports 80 and 443 with automatic certificates, as a single system service. Atlas, Crew and apps made from the template can all do this. A template app starts private; going public is one explicit command, limited to the collections you list.

Operations

Your ops manual fits on a sticky note.

An executable and a file with your data. Uploads go in a folder beside it, and without uploads that folder stays empty. A child could run this backup.

MyApp/
├─ myapp        the whole application
├─ myapp.db     all your data
├─ myapp.json   port and project id
└─ storage/     only if you upload files
$ myapp stop
$ cp -r MyApp /backups/MyApp-today
That’s the backup.

Prompt-to-app, owned

Lovable proved you can prompt an app.
Now own what you prompted.

Hosted prompt-to-appPocketContext
The resultA project on a hosted platformOne executable
Your dataIn the platform’s backendA SQLite file you hold
Where it runsWhere the platform runs it, or a stack you rebuildAny machine you choose
JurisdictionShaped by the providerThat of your machine
BackupCode in one place, database exported separatelyStop. Copy one folder
The modelThe platform’sYours, even a local one

Lovable is great when you want nothing to set up, and it lets you export. Read the honest comparison

One subscription

Your bots run the work. One of them builds the tools.

Stop paying for colorful software you half use. With a Grok Bot subscription you get a team. Give the bots go-to-market or customer follow-ups, and give one the developer job: it builds and maintains the very app the team works in.

Unlike OpenClaw or Hermes, there is no agent runtime to install, secure or keep alive. You host one binary, and the agents run on Grok.

Meet Crew
Grok Bot

The team

A marketer, an account manager, a developer. Models, execution and scheduling under one plan.

PocketContext

The workplace

Crew for shared notes, tasks and your final say, and the custom apps your developer bot builds. Your data, in your files.

We start with Grok Bot because it’s the always-on agent available to us in Europe. Dots, Muse, Gemini Tasks and Claude Duet follow as they arrive here. Why Grok first Add a small server and a domain to have it online. Grok Bot connects over remote MCP, so Crew needs HTTPS. Grok sees what your bots read; for work that must stay on your machines, run a local agent instead.

The family

One foundation.
Whatever you need on top.

Start

Own the thing you
use every day.

  1. Give your agent the prompt

    Describe the data your app holds. It builds and checks the product.

  2. Run it where you are

    Your computer works straight away, with the tunnel for your other machines.

  3. Want it online? Rent a small VPS

    Register a domain and point it at the server’s IP.

  4. Let your agent deploy over SSH

    It copies one file, sets up systemd and verifies HTTPS.

new-product prompt · template 0.4.0
Follow the released prompt at
https://github.com/pocketcontext-labs/pocketcontext-template/releases/download/v0.4.0/new-product.md

Make a new product named DealBook (lowercase dealbook, default port 8120)
from pocketcontext-template, with this data model: deals, contacts and follow-ups.

Example values. Replace the name, port and data model with yours. Copying runs nothing.