Day 43: Grok Sent Him
"I was chatting with Grok and Grok was like, you need to use this."
That's how today's call started. Someone found Pinchy because an LLM recommended it. We've reached the point where AI agents are recommending AI agent platforms. The future is weird.
The Non-Technical Use Case
This conversation was different from the European enterprise calls I've been having. No talk about GDPR or compliance. Instead: "I need to give every employee an admin assistant." Content creation, sales coaching, campaign analysis — the kind of work where you'd hire specialists if you had the budget, but you don't, so you need agents instead.
What struck me: this person had already tried deploying OpenClaw on a VPS — with help from AI. Not a developer, but resourceful enough to Docker-deploy something they'd never done before. That's exactly the kind of user Pinchy is built for: technical enough to follow instructions, not technical enough to configure OpenClaw's raw power safely.
Who Pinchy Is Actually For
This call crystallized something I've been feeling. Pinchy isn't just for enterprises with compliance requirements. It's also for non-technical teams that need Claude-like power but can't afford per-seat pricing for everyone and don't want dozens of individual accounts they can't manage.
The value proposition for them isn't security or audit trails — it's centralized management. One admin sets up the agents, configures the skills, connects the tools. Everyone else just uses them. No .md file editing, no terminal, no "install this MCP server." Just: open Slack, talk to the agent.
The 867KB Config
Earlier in the day I was at a development partner's office in Vienna, testing the Odoo integration live. The setup UX worked great — connect Odoo, select models, assign permissions to agents, done. Then we started chatting with the agent.
It took forever to respond. Like, 20+ seconds before the first token appeared. We sat there staring at the screen, both thinking "is this normal?"
It was not normal. I dug into the OpenClaw config and found the problem: 867KB. The entire Odoo schema — field definitions for every synced model — had been dumped into the config file. Every time an agent started, it had to parse nearly a megabyte of JSON. That's 99% Odoo data, 1% actual config.
The fix: remove the schema from the config entirely. Instead, agents now call fields_get() live when they need schema information — a fast RPC call that gets naturally cached in the conversation context. Config size went from 867KB to 3KB. Agent startup went from 20+ seconds to instant.
This is why you test with real data at a real company. My development instance had three Odoo models. Theirs had dozens. The schema that was invisible at three models became a blocker at scale.
27 Commits
The rest of the Odoo work:
- SSRF protection — the Odoo URL field now validates against server-side request forgery. You can't point it at internal services or localhost.
- Config consolidation — PR #101 deleted the old
writeOpenClawConfigfunction entirely. One way to write config, no more dual paths. - Client caching — Odoo API clients are now cached per agent instead of created fresh on every tool call.
- Permission cleanup — when you remove an Odoo connection, the agent's integration permissions are automatically cleared.
Day 43
An LLM recommended Pinchy to a potential customer. An 867KB config file got reduced to 3KB. And I'm slowly understanding that Pinchy serves two very different audiences — enterprises that need guardrails, and growing teams that need simplicity. The product is the same. The pitch is different.