Audit + handoff layer for AI developers running Claude Code and Codex against the same Obsidian vault. Both agents share `AGENTS.md` / `CLAUDE.md`; the plugin passively observes both and writes a structured Handoff Log so the next session inherits what the last one did — no MCP integration required on either side.
Answers without citations. Threads with no ending. Context scattered across messages you'll never find again. Your vault holds the truth — chat just buries it.
Brain Operator turns chat fragments into a structured, addressable knowledge layer. Same vault. Same model stack. New behavior.
Most Obsidian AI plugins pretend you only run one agent. Brain Operator reads recent sessions from every agent on the machine — alongside your git log — so the timeline reflects what was actually done, not just what was asked.
Each mode is an opinionated operation on the structured vault. Ask retrieves with citations and returns an Answer Trace. Evolve audits and ranks by leverage. Ship classifies and executes. Agent Pulse reads what every agent on the machine actually did. Trace isn't a separate mode — it's the artifact every Ask returns.
The retrieval graph every Ask returns — not a separate mode. Catch drift before it ships. Cited, considered, and unresolved nodes render inline; pin them across follow-ups, fork sideways, or cite into a wiki page.
Question your vault from the Brain Operator modal in Obsidian. Watch retrieval happen. Streamed with citations.
Seven-step audit ranks proposals by leverage. Stale claims, phantom links, goal-adjacent gaps.
Reads the proposal queue. Classifies Level 0 (auto), 1 (draft), 2 (surface).
Read-only timeline. Every agent on the machine, observed.
Four operations. One loop. Continuous improvement. Ingest, structure, audit, propose, ship — then repeat. Your vault gets sharper with every cycle.
These aren't vanity metrics. They're the audit Brain Operator runs every Ask — the holes, the orphans, the stale ground truth, the half-finished drafts.
Your agent has agency. Brain Operator makes that agency observable. No silent rewrites, no hidden side effects. Three rails enforce it.
Default mode for every operator surface. No write capability instantiated unless the user explicitly opts in for a single run.
Every write is constrained by target_paths declared before execution, and validated with git diff --name-only HEAD after.
Local-first by construction. Your subscription, your machine, your filesystem. Nothing leaves the vault unless you commit and push it yourself.
Every Ask returns an Answer Trace — the full retrieval graph rendered inline. What the catalog matched, what was actually read, what got cited, and what didn't resolve. No more wondering whether the model hallucinated a wiki link.
Local-first by construction. Works with your existing tools — Obsidian, Claude, Codex, your terminal of choice. Sixty seconds from empty vault to working brain.
# requirements: Obsidian 1.4+, Claude Code or Codex CLI installed $ cd ~/path/to/your-vault/.obsidian/plugins $ git clone https://github.com/Ege-Deniz/brain-operator brain-ask # in obsidian: Settings → Community plugins → enable "Brain Operator" ✓ first launch scaffolds the brain · Cmd+P "Brain Operator: Home"
# requirements: Obsidian 1.4+, Claude Code or Codex CLI installed $ cd ~/path/to/your-vault/.obsidian/plugins $ git clone https://github.com/Ege-Deniz/brain-operator brain-ask # in obsidian: Settings → Community plugins → enable "Brain Operator" ✓ first launch scaffolds the brain · Cmd+P "Brain Operator: Home"
# requirements: Obsidian 1.4+, Claude Code or Codex CLI installed PS> cd $env:USERPROFILE\path\to\vault\.obsidian\plugins PS> git clone https://github.com/Ege-Deniz/brain-operator brain-ask # in obsidian: Settings → Community plugins → enable "Brain Operator" ✓ first launch scaffolds the brain · Ctrl+P "Brain Operator: Home"
# obsidian community plugins (when published): # Settings → Community plugins → Browse → search "Brain Operator" → Install → Enable # requirements: Claude Code or Codex CLI installed (any one is enough) ✓ technical preview v0.6.1 · pending marketplace review