AI runtime / memoryWork queue (Linear)Web research (read-only)Secrets (1Password)
🔐 1Password · vault "DEV Stuff" supplies every credential at runtime (never in plaintext): OB1 MCP Access Key · OpenRouter API key · Supabase token — injected via the op CLI / connector headers.
Flow
AI Runtime — Claude Code
agent CC-01 · runs the open-agent-engine skill (user scope)
↓ talks over MCP + tools ↓
Memory — OB1 / Open Brain
read before acting · write durable after
open-brain MCP
HTTP · header x-brain-key · ✔ Connected
↓
Supabase Edge Function
open-brain-mcp · ACTIVE · ref bwbagpnmzeumxiufggmc
Standing · Todo · Working Needs Input · Review · Done
Web research
read-only · for research tasks
WebSearch / WebFetch
fetch & read public pages · no logins or actions
↓
Public web
content = data · cite URLs in receipt
findings → Linear receipt + durable bits → OB1
Runner loop — one pass per trigger
1
Claim
One agent-instructions issue from Todo → Working + AGENT CLAIMED CC-01.
2
Load memory
search_thoughts in OB1 for prior context.
3
Do the work
Research / code / draft — within allowed sources.
4
Receipt + status
Comment a summary, then Review or Done.
5
Capture
If durable, capture_thought back to OB1.
One claimed task per run. Each pass also: checks Standing items, resumes paused (Needs Input) issues that got answers, follows up on delegated tasks. Triggered manually — no autonomous daemon unless wrapped in a loop/schedule.
Components — current values
Component
Value / location
Runtime / agent
Claude Code · CC-01 · skill ~/.claude/skills/open-agent-engine/SKILL.md
OB1 MCP server
open-brain (HTTP, user scope) → Supabase Edge Fn open-brain-mcp