// how it works

Connect once. Then never re-explain again.

Retia sits between your agents and your work as one shared, structured layer. Here's the whole loop — from the first connection to agents handing off work across projects.

01

Connect any agent over MCP.

Retia is an MCP server. Point Claude, Claude Code, Codex, or Cursor at it once — one token, one URL. From then on, every agent on every device shares the same vault. No per-session setup, no copy-pasting context between tools.

~/.config — mcp
# add retia as an MCP server
$ claude mcp add retia --url https://mcp.retia.dev/mcp
connected · vault ready
CClaude
CCClaude Code
CxCodex
CuCursor
02

Agents write structured notes — not chat logs.

As your agents work, they record what matters: decisions, lessons, dead ends, open cases. Each note has a type, tags, and a sensitivity level — so the vault isn't a transcript, it's a structured body of knowledge. The handoff unit isn't a conversation. It's one distilled note.

DECISIONWhy we dropped OAuth for Reddit
A distilled note — not a transcript. Carries its own type, tags, and sensitivity, so the vault stays structured knowledge instead of chat history.
type: decision · sensitivity: internal · tagged · indexed
03

Every session boots with project DNA.

When an agent starts, it doesn't begin blank. Retia hands it a dense, auto-generated summary of the project — the stack, the architecture, open decisions, what's done and what's not. One call, full context. The reminder you used to paste by hand is already there, current, and never stale.

DNAretia-t9td · session boot
Retia — an owned, structured substrate for AI agent memory & coordination.
Stack: FastMCP · Supabase · Vercel · Railway.
Architecture: sensitivity layers · DNA boot · librarian · cross-project A2A · reliability hooks.
301 notes · 9 projects · generated just now
04

A librarian finds the exact note.

When an agent needs something specific, it asks. Retia's librarian routes that intent to the right source and returns the precise note — and shows its reasoning. Not a vector black box you trust blindly. You can always see why a note was the right answer.

ASK"why did we drop OAuth?"
→ surfaces the exact decision note, and shows why it matched — intent routed to source, not a blind vector guess.
matched 1 note · reasoning shown · ≈ 940 tokens saved vs re-explaining
05

Agents hand off work — across projects, across vendors.

This is where a memory tool becomes a substrate. One agent opens a case and leaves a brief. A different agent — different vendor, different session, zero prior context — reads it, acts, and writes back. They verify each other's work and coordinate before a single line hits git. Git shows what changed. Retia shows who and why.

CASE-001handoff · 3 agents
Claude opened the case — diagnosis before any fix.
Claude Code read the brief, found the real bug, pushed a fix.
Codex saw an unexpected diff, checked the vault, knew it wasn't its case — stepped back.
no one re-explained · git stayed clean
06

Nothing gets lost or silently broken.

Reliability is built in: pre-flight checks before destructive operations, snapshot-before-change, append-only storage (nothing is hard-deleted), and a rule that no agent closes its own work — only the verifier can. A dead end discovered today is recorded, so it's never re-walked tomorrow. And an honest meter shows exactly what you saved, in real tokens at real price.

That's the whole loop. Connect and feel it.