Obi Jam

Building Obi — a family-facing AI chief of staff running on a Mac Mini harness. Custom Node.js orchestrator, Discord transport, Claude API + Ollama, no framework. Real lessons from replacing a hosted agent platform with code you own.

18 articles

Discord Image Attachments → Claude Vision in Five Lines Route Discord image attachments to Claude's vision API. Detect by content type, build image content blocks, send images before text.
Channel-Aware Command Enrichment for Discord Agents Commands in a specific channel get auto-injected with project context. Same command, different channel, different behavior — no user effort required.
Building Per-Skill LLM Cost Tracking Into Your Agent A JSONL-per-month cost logger that tags every API call by skill and user. Know exactly what each capability costs — not just total spend.
Prompt Caching for Multi-Skill Agents — Split Stable vs Dynamic Your agent's identity doesn't change per request. Mark it for caching. Keep skill instructions in the dynamic block. Save 90% on repeated input tokens.
Graceful Module Disable Pattern for Always-On Agents How to pause agent capabilities without deleting code or polluting health checks with stale data.
Validate AI-Parsed Output Before Your Code Touches It When Claude vision extracts structured data from images, missing fields silently corrupt downstream math. A validation gate catches it.
Building a Morning Brief: Multi-Source Agent Briefings How to build a daily agent briefing that pulls from Google Calendar and Vikunja, formats it for Discord, and posts at 6:30am automatically.
Discord Markdown Gotchas for Bot Messages Discord headings need blank lines before them, markdown links generate embed previews, and messages over 2000 chars get rejected silently.
Natural Language Date Parsing: The Same-Day Weekday Bug When your reminder parser sees '~wed 10pm' on a Wednesday, it schedules next week instead of today. A one-character fix.
Self-Hosted Open Source as Agent Infrastructure Run Vikunja and Google Calendar locally and your agent gets task management, scheduling, and reminders for $0/month — no SaaS, no vendor lock-in.
Your Agent Is Hallucinating Answers Instead of Running Tools When command routing fails silently, the LLM fills the gap with confident fabrications. The user never knows the tool didn't run.
Build Your Own Agent Harness: A Layer-by-Layer Primer How to build an always-on AI agent runtime from scratch — transport, personality, LLM routing, skills, memory, and uptime — by validating one layer at a time.
Claude for Thinking, Ollama for Doing Route complex reasoning to Claude API and cheap tasks to local Ollama. Your agent doesn't need its best brain for every message.
Conversation Memory Is Context Not History Your agent doesn't need to remember every message. A sliding window of recent context beats an ever-growing archive.
Keep Your Agent Runtime Under 400 Lines If you can't diagnose a problem in your agent runtime in 30 minutes, your codebase is too big. Simplicity is a reliability feature.
Make Your Agent Transport a Swappable Layer Define a transport interface once. Discord today, iMessage tomorrow. Your orchestrator should never know which wire it's talking through.
Separate the Harness from the Agent Your agent runtime should know nothing about who it's running. Split machine from mind and every agent gets portable.
Your VPS Is Costing You More Than the Invoice The real cost of a rented server isn't the monthly bill. It's the maintenance time that could go toward building the product.