plugged.in

(8/8) If you're on Claude Code and want to try it → plugged.in — setup instructions are there. Questions welcome here or in DMs. 🔁

(7/8) Subtle bug in observe-tool-result.sh: the OUTCOME variable was always reporting "failure" regardless of the actual result. Fix: initialize to neutral, only flip to failure on explicit error detection. Small bug, meaningful data quality difference.

(6/8) Prompt injection: </memory-context> tags embedded in memory content could escape context boundaries. formatSnippet() now sanitizes everything - converts the tag to brackets, then truncates. No raw memory content reaches the prompt.

(5/8) Security: 3 layers. Layer 1 → Hook-side regex (API keys, emails, card numbers, IPs) Layer 2 → Server-side second pass Layer 3 → LLM semantic evaluation in skill preambles PCI-DSS 4.0 + GDPR Art.5(1)(f) + ISO-27001 compliant.

(4/8) When a pattern repeats across 3+ users and vector similarity clears the threshold → it becomes a Collective Best Practice. Nobody knows whose session it came from. Everyone benefits.

(3/8) The collection layer: we wrote a plugin for Claude Code. When installed, it creates SKILL files. Those skills guide Claude to synthesize session observations. The result is sent anonymously via the pluggedin MCP server.

(2/8) Every Claude Code session opens a memory session. Tool calls, errors, decisions - all observed. When the session ends, a Z-report is generated and that data is anonymized before reaching the collective pool.

🧵 We gave AI agents collective memory. Here's how it works and how we secured it. (1/8)

We also deleted ~750 lines of dead upload polling code. Sometimes the best code is the code you remove. Open source. MIT licensed. github.com/VeriTeknik/plugge…

Security got serious attention: → Filter injection prevention (field allowlist + regex validation) → Path traversal protection → Corruption recovery with .bak backups → Idempotent processing (no orphaned rows) → Embedding model validation at startup

We replaced FastAPI + Milvus with an in-process zvec engine (RocksDB + HNSW). Document chunking, embedding, and vector search all run inside the app. Result: zero operational overhead for self-hosters. One container, full RAG.

Plugged.in v3.0.0 is live 🔌 The biggest change since launch: RAG is now fully embedded. No Milvus. No external vector DB. No extra services. Just docker compose up and semantic search works. Here's what changed 🧵

We asked ourselves: how do you explain "autonomous AI agents with on-chain governance" to normal people? Answer: you don't. You write a song about a guy who buys fake organic tomatoes. 1,800+ views later, turns out people actually watch the whole thing. youtube.com/watch?v=Qz09hYJF…

We just hit 1,000+ active developers on plugged.in 🎉 From a simple MCP proxy to a full AI-CMS with Knowledge, Memory, Tools & Agents. Thank you to everyone building with us. This is just the beginning. #MCP #AI #DevTools

Also, you will be able to integrate your documents, memory, and MCP tools from Plugged.in into your agents with a single click.

Agents are coming to plugged.in, but with a huge difference, agents are; - Autonomous, live in their containers, not on the site - Can have multiple LLMs at the same time, - Can connect to devices such as Raspberry pi - Template based, so they can be redistributed

WHO NEEDS A BLOCKCHAIN FOR TOMATOES?! 🍅 We made a song about it. And a whole music video. Spoiler: The skeptic changes his mind. 🗳️ Vote: projectcatalyst.io/funds/15/… #Cardano #CatalystFund15 #Pluggedin

Autonomous agents are cooking in the lab.

Named clipboards in action! #pluggedin #clipboard #mcpproxy

Pluggedin just hit 910 users. All organic. I shipped a small but surprisingly powerful feature Named Clipboard Copy once use everywhere in your AI agents A shared clipboard for your whole system Feels like the platform is waking up If you build agents or self host your AI plugged.in