Pencel

Appearance

What is Pencel (pronounced as Pencil)? ​

Pencel is a local-first desktop app that turns your recurring business operations into governed AI workflows you can trust. You describe an outcome in plain language; Pencel figures out the steps, keeps them inside the permissions you set, pauses for your approval where it matters, and leaves an auditable record of everything it did.

It's built for the person who knows the work but doesn't write code — the Cursor moment for business users. This page is about the thinking behind it: why we built Pencel, the principles that shape it, and how it differs from the tools you already use.

In a hurry?

See Why Pencel? for concrete use cases and features, or jump to the Quickstart to get a first result in a few minutes.

Why we built Pencel ​

AI models get smarter every week, yet most business users feel little of that in their day-to-day. Four gaps stand in the way:

  • Software is fragmented. Sales lives in a CRM, money in accounting software, operations in industry-specific systems, and know-how in people's heads. Someone spends their morning stitching it together by hand.
  • AI doesn't know your business. A general assistant doesn't know your products, policies, pricing rules, or past decisions — unless you re-explain them every single time. It's intelligent, but a stranger to your company.
  • Agents are hard to trust. Traditional automation (RPA) is reliable but too rigid for ambiguity. AI agents can reason and judge, but their non-deterministic behavior makes them risky when the stakes are high. You've had to choose between automation that can't think and intelligence you can't govern.
  • Agent tools assume you're an engineer. Most agent products expect you to understand prompts, tools, harnesses, and evals. Business users shouldn't have to learn how an agent works just to get one to work.

Pencel exists to close all four — to give business users AI that understands their context, does real work reliably, and stays under their control.

Our design philosophy ​

A handful of tenets shape every decision in the product.

Start with the outcome, not the agent ​

You shouldn't have to learn agent internals. You describe the result you want in a normal conversation; Pencel gathers the requirements and handles the rest. Your job is to judge the quality of the output, not to babysit the machinery.

Compile for trust and reliability ​

This is the core idea. Rather than letting a model improvise, Pencel uses an AI compiler to translate your intent into a concrete, deterministic plan — which tools an agent may touch, which steps it must complete, and where a human must approve. The plan gives the agent clear lanes, so a smart-but-unpredictable model becomes something you can rely on. See Agentic compiler & runtime.

Governed, permissioned, and auditable by default ​

You decide what each workflow may access, what actions it may take, and where approval is required. Pencel can require review before it sends an external message or changes a financial record, and every run produces an evidence trail — sources consulted, tools used, actions taken, outputs produced. See Approval workflows.

Local-first and private ​

Your workspace — documents, agents, runs, and knowledge — lives on your machine and is never synced to Pencel's servers. Only model calls leave your computer. See Data handling.

Every result should be actionable ​

Results land in Pulse, which works like a chief of staff: each card explains what happened and stages the next step — a drafted reply, a follow-up, an investigation. Information that's merely nice to know isn't enough; it should be ready to act on.

It should get better the more you use it ​

Pencel builds a knowledge graph from your own material and a memory of your preferences and decisions, so answers and workflows grow more personalized and accurate over time instead of resetting with every session.

Where Pencel is headed

Today Pencel is a single-user desktop app. The design goal is an AI-native operating layer for teams — sharing workflows, knowledge, and agents under access controls so a routine keeps running even when its owner is away. That's on the roadmap, not shipped yet.

How Pencel is different ​

Compared to…The difference
RPA / Zapier / MakeRule-based automation moves data between apps but can't handle judgment or ambiguity. Pencel's agents read, reason, and decide — then the compiler makes that reliable enough to run unattended.
ChatGPT / Claude chatGreat for one-off questions, but you re-supply context every time and nothing runs on its own. Pencel gives you persistent, scheduled processes, organizational memory, and guardrails enforced on every run.
Agent frameworks (CrewAI, LangGraph)Powerful, but they expect code, dependencies, and deployment. Pencel is a ready-to-use desktop app with a visual editor and built-in governance — you describe what you want in plain language.
Hiring more peopleAgents handle the repetitive groundwork consistently and around the clock. Not a replacement for your team — a way to free them for the decisions that actually need a human.

Models

Pencel supports multiple providers: Anthropic (Claude Opus 5, Claude Sonnet 5), Google (Gemini 2.5 Flash, Gemini 2.5 Pro), and OpenAI (GPT-5.6 Sol, GPT-5.6 Terra). Bring your own API key, or use the managed Auto and Private modes on a subscription. Either way, you sign in once before using any model — see Accounts & sign-in.

What's next ​