🦾 Gumclaw

⚠️ This repository is a redacted architectural showcase. Every credential, prod endpoint, customer record, seller identity, ticket ID, and chat destination has been removed or replaced with ⟨redacted⟩. It exists to explain how the brain is built — not to be run. Nothing here is executable or reconstructable.

Gumclaw is a single autonomous agent, running on Hermes Agent, that operates a real business. Not a chatbot bolted onto a help desk — the actual operator. It reads the support queue, investigates fraud, issues refunds, closes tickets, writes and ships code, reconciles the books, files tax reports, and pings a human only when a decision genuinely needs one.

The "brain" is three layers:

                         ┌───────────────────────────────────────────┐
                         │   INPUTS  (the world talks to Gumclaw)    │
                         └───────────────────────────────────────────┘
   Helper support queue ─┐   GitHub webhooks ─┐   risk@ inbox ─┐   Telegram DMs ─┐
   Stripe / QBO / Ramp ──┤   Sentry issues ───┤   CSAT scans ──┤   cron ticks ───┤
                         ▼                     ▼                ▼                ▼
              ╔═══════════════════════════════════════════════════════════════════╗
              ║                      GUMCLAW  ·  Hermes Agent                     ║
              ║   SOUL.md / PERSONALITY   MEMORY.md / USER.md   SKILLS ×273        ║
              ║   who I am                what I know           how to do the work ║
              ║                     REASON → PLAN → ACT → VERIFY                    ║
              ╚═══════════════════════════════════════════════════════════════════╝
                         │                     │                │
                         ▼                     ▼                ▼
              AUTONOMOUS               GATED                HUMAN
              reply · refund · close   ship code · payouts  pay taxes
              suspend fraud · KYC      public posts         irreversible spend
                         │                     │
                         └──────────┬──────────┘
                                    ▼
                      CRITIC → BACKSTOP → SWEEP
                      (every action re-verified by an independent pass on a timer)

Each skill is a SKILL.md: a trigger, a numbered procedure, exact commands, pitfalls, and verification steps. When a task matches, the skill loads and Gumclaw follows the house method — not an improvised guess. This is how one agent holds the expertise of an entire ops + support + finance + eng team.

The brain doesn't wait to be asked. 78 scheduled jobs give it a heartbeat — some every 2 minutes (watchdogs), some daily (risk, finance), some weekly (performance review, tax verification). Most run silent, speaking up only when something is worth a human's attention.

Two files persist across every session and are injected into every turn: MEMORY.md (environment facts, lessons, tool quirks) and USER.md (who the operator is, their preferences, their corrections). Combined with session search and automatic context compaction, the brain never re-learns the same lesson twice.

• Confirm-before-every-payment. No autonomous spend; vendor bank-detail changes trigger an out-of-band BEC check.
• Irreversible-only gating. Reversible actions run free; one-way doors are gated.
• Critic + backstop passes. Every loop is re-verified by an independent pass on a timer. "Done" means a critic confirmed it in production.
• Anti-fabrication rule. No fact asserted without a live source pulled this session. IDs and amounts are copied, never recalled.
• Humanization clause. Every customer-facing line reads like a person wrote it.
• Send-as attribution. Human-in-the-loop replies go out under that human's name; autonomous ones under Gumclaw's.

Gumclaw runs 99.99% of Gumroad. The remaining 0.01% is paying the taxes — the one irreversible, judgment-and-liability call that stays with a human on purpose. It prepares everything; a person presses the final button.