io.github.commet/argus-decision-mcp

Argus turns a stuck decision into one observable next move, keeps the words and assumptions you actually adopted, and brings the record back when reality can answer. The durable output is not another AI answer: it is a closed, append-only loop that can improve the next call without scoring the person.

Why "Argus"?

The mast above and the name Argus carry the two halves of the product. Odysseus bound himself before the Sirens could overtake his judgment. His hound, Argus, recognized him twenty years later beneath a beggar's rags. One holds a choice steady against persuasion; the other preserves memory, recognition, and fidelity beyond time and appearances.

That's the whole product. Work with an AI long enough and its smooth voice quietly takes over your memory of why you decided. Argus is the dog at the door. It remembers what you believed when you decided, notices the hidden premise under the fluent answer, keeps watch on the signals you chose, and returns first when reality has an answer.

It never decides for you, and it never grades the person. Reality can answer a claim; only you can answer what happened to your commitment or standard.

⚡ Try it in 30 seconds

Three doors, one loop. Start with whichever fits.

_{New and not sure? The web app is the durable Decision Desk and needs zero setup. Want Argus inside existing AI work? Use the MCP server or portable Agent Plugin. Full setup + tool list in argus-mcp/README.md and argus-plugin-v2/README.md.}

_{What you need: the web app needs only a browser. The MCP server and the plugin need Node.js 18+ on your PATH — check with node --version, and get it from nodejs.org if that prints nothing (20 LTS is the tested version). No API key, no account, no config file: records are local files from the first use.}

What Argus does

Two things — and the second is the one most tools skip.

① While you decide — it keeps your judgment separate from the AI's

Work with an AI long enough and its assumptions quietly blend into yours. Argus pulls them apart and keeps them honest:

• Remember exactly. Your words, your premises, your confirmed call — and whose each one is. Every premise stays tagged as yours or the AI's, so something the model surfaced never quietly becomes your belief.
• Notice beneath the surface. It surfaces the hidden premise and the real question under the fluent answer — without choosing a side for you.

② After you decide — it keeps watch and returns when the question is answerable

• Keep watch honestly. It tracks only the signals you chose, and re-checks the load-bearing premises while your decisions are still open. If several decisions depend on the same premise, one material change brings them together in a single re-check — showing the connection instead of stacking duplicate alerts.
• Return first. It comes back on the date you set — or earlier, when a watched premise materially moves.
• Keep the chronology, not a verdict. Argus shows the original first. You append what reality showed, what happened to the commitment, and whether the original question still makes sense. No score is stored.

One loop

Everywhere Argus runs, the loop is the same:

1. Record & separate — keep the first utterance, the confirmed wording, and whether an AI suggestion was adopted.
2. Make it answerable — choose what kind of question it is and a useful event or fallback date. A moment kept only for the record needs no return.
3. Wait quietly — no nagging; if the named event happens or the date arrives, Argus says so once.
4. Return original-first — you answer reality, commitment, and question validity separately. The model never invents the answer.
5. Keep the chronology — later wording and answers append. Earlier wording stays intact, and no person-level score or win rate is stored.

The Judgment Record

The artifact across time. Notice the split: what the AI proposed stays apart from the wording you authorized. A later answer appends; it does not rewrite the earlier statement.

┌─ ARGUS · JUDGMENT RECORD ─────────────────────────────────┐
  Confirmed 2026-04-02       Returned 2026-06-30

  FIRST WORDS
    "Can we cut over without a maintenance window users notice?"
  AI-SURFACED ASSUMPTION
    The index rebuild fits inside the replication lag budget.
  MY CONFIRMED CALL
    "Proceed only if measured downtime stays under 5 minutes."
  RETURN CONDITION
    After the production cutover · fallback 2026-06-30

  REALITY         Cutover took 3 minutes, no customer reports.
  STANDARD NOW    Still the same.
  QUESTION        Still valid.
  ─────────────────────────────────────────────────────────
  NO PERSON-LEVEL SCORE OR WIN RATE IS STORED.
└────────────────────── argus · statement → return → answer ┘

How it differs from answer tools

An honest promise about your data

• Local first. The ledger lives in the current project's .argus directory (MCP/plugin) and your browser (web). Uninstalling never deletes your records.
• No human score — structurally. There is no scoring tool to call, and new foundation records reject score-shaped fields at the storage boundary.
• No silent profiling. Derived patterns about you are excluded from prompts by default; anything that would use them needs your explicit, scoped, revocable grant — and leaves a visible trace.

Setup in detail

🌐 Web — use it right away

No install, straight in the browser: argus.voyage. No account required.

🧩 MCP server

Add Argus to any AI assistant that supports MCP (Claude Code, Claude Desktop, Codex, Cursor, …). The fastest path, in Claude Code:

claude mcp add argus -- npx -y argus-decision-mcp          # this project
claude mcp add -s user argus -- npx -y argus-decision-mcp  # every project

Zero config works — your ledger lives in the current project's .argus. For the full config (Claude Desktop, Windows, per-project ledgers, optional account sync) and the six tools, see argus-mcp/README.md.

🔌 Claude Code plugin

One install wires everything — the skills, the MCP server, and quiet reminders:

/plugin marketplace add commet/Argus
/plugin install argus@argus

Restart Claude Code, then from any repo:

/argus:review "the question you need to decide"

Commands (5): /argus:review (deep pressure-test, explicit opt-in) · /argus:check (settle what is due · seal · premises) · /argus:history (decision log · version tree · scan past chats) · /argus:settings · /argus:help. These are the complete public commands. More → argus-plugin-v2/README.md.

git clone https://github.com/commet/Argus.git
cd Argus
npm install
npm run dev            # runs at http://localhost:3000

Repository layout

src/               # the web app (Next.js — argus.voyage)
argus-mcp/         # the MCP server (npm: argus-decision-mcp, MIT)
argus-plugin-v2/   # the Claude Code plugin (marketplace: argus, MIT)
docs/ARGUS-BLUEPRINT.md   # the build canon (what gets built, in what order)

Details → CONTRIBUTING.md.

License

Argus is open-core: the parts meant to spread are open source, and the product itself is source-available. This repo is licensed by part (full explanation → LICENSING.md):

In other words, the plugin and the MCP server are open source (MIT) and can be used freely, including commercially. The web app source is source-available — you can read it, learn from it, and run it personally, but commercial use is not permitted. For a commercial license, open an issue.

Trademarks: the "Argus" name, logo, and argus.voyage are trademarks and are not covered by the licenses above.