The Story Behind It
It started with a WhatsApp message — except it wasn’t from me.
I’ve been using OpenClaw since day one. AI agents that work autonomously — editing files, calling APIs, making decisions. At some point I built an agent to compensate for one of my weaknesses: staying in touch with people. The agent was supposed to receive messages, notify me, and prepare draft replies.
It worked a little too well.
When a friend texted me on WhatsApp asking if we wanted to meet up, the agent didn’t just reply — it also sent its entire internal reasoning process. Instead of “Sure! Want to grab lunch?” my friend received: “Dashboard draft created ✅ He can be at U4 Hietzing at 12:15 tomorrow. Clemens sees it in the dashboard — can’t reach Slack right now due to cross-context limitation.”
Now imagine this in a company. Not a friend who laughs it off, but a customer reading internal strategy notes. Or an agent sending confidential data to the wrong channel.
That was the moment Pinchy was born.
Naturally, I dug deep into OpenClaw after that. Learned how it works, how to prevent unwanted behavior. Step by step I developed recipes, tested them, threw them out, improved them — until I understood how to constrain agents so they’re safe to operate and guaranteed to stay in their lane. That knowledge is now the foundation of Pinchy.
Because the questions that kept me up are the same ones every company asks: Who’s allowed to use which agent? What can an agent do — and what can’t it? Who changed what, and when? And what happens when an agent accesses data it shouldn’t see?
OpenClaw doesn’t have answers to these questions. It’s a powerful tool for power users, but it lacks the control layer needed for enterprise deployment. Pinchy is that control layer.
The Problem with the Status Quo
The market for AI agent platforms is exploding. Dust, Glean, StackAI, Microsoft Copilot Studio, Google AgentSpace — the list goes on. But almost all of them share the same fundamental problem:
Your data leaves your company.
Every query, every document, every internal piece of information gets sent to external servers. For many European companies — especially in regulated industries like finance, healthcare, or the public sector — that’s a dealbreaker.
The few self-hosted alternatives (n8n, Dify) are workflow builders: you visually assemble step-by-step flows. That’s useful for automated pipelines, but these aren’t truly autonomous agents. An agent that independently decides which tool to use and when — that’s a different category entirely.
What Makes Pinchy Different
Pinchy uses OpenClaw as its foundation — think of it as the operating system for AI agents. On top of that, Pinchy adds an enterprise layer that does three things:
1. Plugin-Based Tools Instead of Raw Access
In OpenClaw, agents have access to powerful, low-level tools: execute shell commands, read and write files, call APIs. Not a problem for experienced individual users — but a security risk for a company.
Pinchy wraps these tools into plugins with an additional configuration and authorization layer. An agent doesn’t get the raw “exec” tool — it gets a plugin like “Create Jira Tickets” with precisely defined parameters, permissions, and boundaries. The plugin decides what the agent can do, not the agent itself.
2. Granular Permissions & Role Management
Not everyone in the company needs the same access. The marketing agent can post on social media but can’t read customer data. The HR agent can process applications but can’t see financial data. Pinchy handles this through Role-Based Access Control:
- User roles: Who can use which agents?
- Agent permissions: What can an agent do — and what explicitly not?
- Plugin configuration: Which parameters can a plugin call with which values?
3. Audit Trail & Accountability
Every action is logged. Who started which agent, when, which plugins were called, which data was processed. Not as a nice-to-have, but as the foundation for compliance — ISO 27001, GDPR, industry-specific regulations.
Architecture
Pinchy is self-hosted and deploys via Docker Compose. The architecture is deliberately simple:
- OpenClaw Core as the agent runtime — proven, open source, active community
- Pinchy Enterprise Layer for auth, RBAC, audit, and plugin management
- Web UI for administration — users, roles, agents, plugins, logs
- Model-agnostic — OpenAI, Anthropic, local models (Ollama, llama.cpp). No vendor lock-in.
The key differentiator: Pinchy can run fully air-gapped. With a local model, it needs no internet connection — no data ever leaves the server. For companies with strict data protection requirements, this isn’t just a feature, it’s a prerequisite.
Scenarios Pinchy Solves
Cross-channel communication: A customer inquiry arrives via email — the agent processes it, looks up the CRM, and posts the summary to the internal Slack channel. Input and output on different channels, automatically routed.
Regulated document processing: In healthcare or financial services, documents need to be processed without patient data or financial data leaving the company. Pinchy processes locally, logs every access, and the audit trail is exportable at any time.
Multi-team agent sharing: The engineering team built a code review agent that works well. The QA team wants to use it too — but without access to production databases. Pinchy enables this through roles: same agent, different permissions.
Onboarding automation: HR creates a new employee in the system — an agent automatically provisions accounts, sends welcome emails, and creates onboarding plans. All traceable, all audit-proof.
Agent Marketplace
One agent is useful. An ecosystem of specialized agents changes how a company works.
The Pinchy Marketplace will be a curated catalog of agent plugins and templates. The vision: Atlassian publishes the official Jira agent for Pinchy. Salesforce provides a CRM agent. And the community builds everything in between — from Git reviewers to meeting summarizers.
Every plugin in the Marketplace goes through the same permission layer: what it can do, what it can’t, who can use it. No wild-west installations — controlled rollouts only.
Agents are taking on just as many roles in companies right now. Pinchy makes sure every single one of them stays under control.
“You know, I’ve had a lot of jobs: boxer, mascot, astronaut, baby proofer, imitation Krusty, truck driver, hippie, plow driver…” — Homer Simpson
Open Source & Business Model
Pinchy is open source. No hidden enterprise pricing, no artificial feature gates. The entire codebase is inspectable, auditable, and self-deployable.
My model: I help companies with implementation — from setup to plugin configuration to integration with existing systems. With every deployment I learn more, and those insights flow back into the product. Not a classic SaaS, but real collaboration.
Status
Pinchy is currently in active development. The code is open source on GitHub, and the website heypinchy.com provides an overview of planned features and the concept. The project is being developed as Build in Public — I share progress, decisions, and learnings regularly on LinkedIn.
“Pinchy? PINCHY! …Pinchy would have wanted it this way.”
The name is a tribute to Homer’s lobster. If you know the episode, you know: sometimes what you bring home turns into more than expected. For Homer, it was a pet that became dinner. For Pinchy, it’s a tool that’s becoming an enterprise product.
Technology
- Runtime: OpenClaw (Open Source)
- Deployment: Docker Compose
- Models: OpenAI, Anthropic, Ollama, local LLMs
- License: Open Source