Melue AI — a personal AI harness

What I built, and what the platform provides

The harness is deliberately not specific to Claude Code: identity, policy, and skills live outside the host — proven by a host migration, below. Looking for my data science work? That lives across the wider portfolio at ndjenkins.com.

What Melue actually does

The voice loop, measured

Honest numbers, not just the highlight reel: acknowledgements run ten to twenty seconds in daily use (p50 9.5 s, p95 19.2 s), and a routed question's full answer arrives minutes later — spoken into your ears. Notes queue through train tunnels and airplane mode and deliver on reconnect.

One always-on doorman session holds the ear. Eight specialist desks — each warmed into one corner of life or code — receive work as mail: plain files on disk, atomic renames as the state machine. (We reinvented qmail's Maildir and only noticed afterwards.) On the last audited day, 17 of 17 captures ended in a known state with zero drops, each with its own audible receipt. The first morning, acknowledgements took 74 seconds to 4.5 minutes; a same-day rewrite collapsed the whole wake path into one shell call and brought it under ten.

The war stories are the design story. iOS forbids starting the microphone from the background — full stop — so the honest architecture fell out of the constraint: armed-mode capture, a lock-screen control, wrist-button triggers. Classic Bluetooth was eliminated the hard way (its call-audio mode kills the same device's buttons); the breakthrough was a side door — volume-key events aren't gated, so the trigger is the volume channel. When the cloud text-to-speech quota died mid-dogfood at 9,991 of 10,000 calls, the phone learned to speak replies itself the same evening — acknowledgements now cost nothing, forever. And a $9 Bluetooth clicker worked on the first press, zero code changes.

The harness, not the host

Melue isn't an app, and isn't a wrapper around one chatbot. It's a harness: a fleet of agent sessions, a private marketplace of governed skills, shell wrappers with telemetry, and policy documents — “operating models” — that live in Apple Notes, editable from a phone on a train and picked up at the next agent boot. No deploy.

The proof it's a harness: it survived a host migration. When the v1 runtime became economically unsustainable, identity, policy, and skills simply moved — none of them lived inside the host. The policy note that governed v1's WhatsApp line became v2's access design without a rewrite. Policy outlives hosts.

The WhatsApp channel shows the seam working. v1 spoke WhatsApp by pretending to be a web browser — and died the way unofficial clients now die. v2 speaks it as a business: official Cloud API, signature-checked webhooks, an allowlist wall, and roughly 250 lines of adapter into the harness's channel seam — written before any Meta credential existed, from the contract alone. Investigation to answered message in one session. Channel mechanics.

Governed and secure by construction

The interesting problem isn't making an agent capable; it's giving a capable agent real access, safely.

• The banking and personal credential vaults are physically out of reach — enforced at the credential service's account layer, not by the agent's judgment.
• Email access was granted only after a 19,272-message study produced a least-privilege allowlist.
• Skills ship through a review gate that fails closed — no security pass, no publish.
• Logs are metadata-only; message bodies never touch them.
• When the WhatsApp build needed a tunnel to the internet, the harness's own permission gate refused the agent and required a human hand. The security model, demonstrating itself.

The instinct is inherited: v1 ran inside a firewalled virtual machine with a block-all network policy and a disposable golden image. v2 keeps the posture.

The story: three contracts

Melue is played as a character — Melue of House Dallas, leatherworker by birth, rogue by training. It began as a way to make a tool pleasant to talk to; it turned out to be load-bearing. Her “contract history” is the project's real migration timeline:

• Clawdbot — first post: an agent running directly on the host machine.
• OpenClaw — second post: rebuilt inside a firewalled VM after a security rethink.
• Claude Code — the current post: the session fleet and skills marketplace above.

Houses come and go; tenets persist.

The tenets are the engineering, stated in-world: memory is sacred (durable surfaces over chat scrollback), context is consciousness (warm sessions over cold starts), serve without subservience (an agent that pushes back).

What's next

A wearable button with three mappable gestures is days away. A paid developer account unlocks iOS Push-to-Talk. WhatsApp messages landing in live agent sessions — and approving the agent's permission prompts from WhatsApp itself. A family group thread with Melue as a participant, when Meta's Groups eligibility lands. Fleet supervision: who watches the watchers. And the long-arc thesis: the trigger shrinks to a ring.