ScoutAn autonomous agent for Claude Code
There's always more happening than you can keep up with.
So I built an agent that keeps up for me — reads everything, checks its own work, and reaches back only when it's worth it.
Reply to Priya — she's asked twice. why ›
The migration doc she needs is already in your drafts.
Renewal call may have moved to 3:00. why ›
Only the calendar says so — treat as tentative.
Did PROJ-214 ship? Sources disagree. why ›
Part I — Trust by construction
No source gets taken at its word.
Cross-checked, not collected
It reconciles your tools before it tells you anything.
A calendar invite checked against the transcript; a ticket against its PR. And what you actually did beats what a note says you'll do.
Confidence on every line
You always know how much to trust it.
Every claim carries its evidence level. A one-source claim is never quietly promoted to fact.
Disagreement is the signal
When tools conflict, it shows both sides.
It won't pick a winner. That contradiction is usually the most important thing in your day — and what every other tool hides.
Part II — It runs & improves itself
The part nothing else does. Six loops it runs on its own.
System 01 · self-knowledge
The mistake audit
It keeps a model of its own mistakes.
Every time it's wrong and you flag it, it logs a numbered pattern — the error, the root cause, the fix — and tracks whether that fix actually held. A "fixed" one that recurs flips to regression, so nothing rots quietly. Over a hundred patterns in, the same class of error keeps getting rarer.
System 02 · self-modification
Dreaming proposals
It rewrites its own instructions — in the open.
Each night it decides what to change about itself. Small, additive fixes it just applies; anything bigger it proposes and auto-applies in a few days unless you reject it. Every change is a separate git commit you can read and revert — transparency instead of a gate on every step.
System 03 · self-direction
The wishlist
It drives its own roadmap.
It keeps a running wishlist of improvements and works it down on its own — picking an item, building it across however many nights it takes, and shipping it. The Mac app, the terminal UI, the research mode: features it built for itself. You're not the only one filing requests.
System 04 · self-expansion
The research queue
It goes and finds what it's missing.
Between briefings it scores which parts of your knowledge base have gone stale or thin, rotates across your projects so nothing's forgotten, and digs — web, docs, your tools — to fill the gaps. Drop a note in the vault and it jumps to the top of the queue.
System 05 · asking for help
Enrichment questions
It asks you only what it can't find.
Some facts no tool can see — who really owns a decision, why something stalled. So it occasionally surfaces one targeted, low-stakes question, always with a link and a one-line gist so you can answer in seconds. Your answer routes straight into the graph, and it never asks twice.
System 06 · knowing its limits
The review queue
It refuses to guess about people.
The most dangerous edit a knowledge system can make is mistaking one person for another. Anything it isn't sure of — especially identities — it holds for you to confirm or reject, instead of silently corrupting the graph.
A full cold start. It reads the entire knowledge base, queries every connected tool, and cross-checks every finding against the others before it writes a single thing — then hands you today's short, trustworthy list.
Light delta scans. It looks only at what changed since the last run — a new reply, a shipped ticket, a meeting that moved — and reconciles each open item against the fresh data, flagging anything that's gone stale or now contradicts.
It goes outward. It scores which parts of the knowledge base have gone stale or thin, rotates across your projects so none gets ignored, and deep-dives with web search and docs to fill the gaps it found — folding what it learns back into the graph.
Self-improvement, while you're offline. It processes the day's feedback, logs its own mistakes, prunes what no longer matters, and rewrites its own instructions so tomorrow's runs are a little sharper than today's.
You wake up already briefed.
While you slept, Scout read everything and cross-checked it. The day starts with one short, trustworthy list — not a pile of unread tabs.
It keeps pace as the day moves.
Light consolidation runs reconcile what changed since the morning — new replies, shipped tickets, a meeting that moved.
It goes looking, on its own.
When there's budget and something's gone stale, it researches — rotating across your projects so nothing is quietly forgotten.
It gets better while you're offline.
Dreaming processes the day's feedback, prunes what no longer matters, and rewrites its own instructions for tomorrow.
These four run by default — and you can retime them, add your own, or let Scout invent its own.
Part III — A memory you can read
A living, validated knowledge graph — not a notes pile.
The graph
Typed people, projects, tasks — and the links between them.
Densely interconnected on purpose: corroboration needs more than one edge to lean on. Validated against a formal ontology on every run.
Git as memory
Every run is a commit. The history is its memory.
Trace exactly what it changed, when, and why — and undo any of it. 800+ committed sessions and counting.
Part IV — Built for anyone, owned by you
It runs on your machine, your tools, your files.
Endless connectors
Connect anything. The list is never closed.
It auto-discovers whatever you've wired into Claude — official, custom, or community. Adds no new cloud; there's no Scout server.
Yours to keep
Plain markdown, in a folder you own.
Human-readable files with [[wikilinks]] — open it, grep it, walk away with it. Not locked to one model or harness.
Install
Run it yourself in under five minutes.
Two commands in Claude Code, then /scout-setup detects what you've connected and scaffolds your vault.
# in Claude Code:
/plugin marketplace add Raven-Scout/scout-plugin
/plugin install scout@scout-plugin
# then create your vault:
/scout-setup