Cole Tebou

§ What it sees

A daemon on each machine watches the coding agents running there: Claude Code sessions, Codex threads, Grok bots and an OpenClaw fleet manager. Together they sort every agent on every host into three lanes, so the one waiting on you is never buried under the ones still working.

§ Where you drive it

The same fleet is reachable from a web UI, a terminal UI, an MCP server that other agents can call, Mac, iPhone and watch apps, and browser extensions. You can answer an agent, message it as if you typed into its pane, or launch a new one on another machine.

• One daemon per machine, zero npm dependencies.
• No open ports.
• Apache-2.0, with light and dark screenshots of every surface in the README.

§ The problem

Every coding agent starts each session knowing nothing it learned yesterday, and each tool keeps its own private notes. A fleet of them re-learns the same facts and contradicts itself.

§ The memory

One self-hosted service that Claude Code, Codex, ChatGPT and any MCP client read and write. Search mixes keywords and vectors; a fact that changes is superseded rather than duplicated, so it keeps its history; every agent has its own key; and everything read back carries where it came from.

• SQLite FTS5 with sqlite-vec for search.
• A write gate that refuses anything shaped like a secret.
• MIT licensed.

§ What it carries

The shell, tmux and git setup every machine shares, plus the agent layer on top: hooks for Claude Code and Codex, a marketplace of skills, worktrees for parallel sessions, chains of session handoffs, and the memory the agents keep between them.

§ One attention list

A message of the day for a fleet: the machine's vitals and services, the AI agents and how much of their plan is left, Docker, open pull requests and cloud status, ranked into a single list of what needs you. It renders from cache, so the login never waits on the network.

§ What it does

It reads your Grok Bots, their routines and their skills, and writes each one to git as plain files. It never writes back: unofficial, and read-only on purpose.

§ How it works

A remote MCP server gives the bot an SMS line, and a skill teaches it to use it. Text the number from any phone and the bot answers in the thread.

§ Product thesis

Model choice should feel like a capability, not a configuration chore. Badland exposes fast and deep reasoning, direct-provider paths, image and voice modes, multi-model councils, files and web research through one consistent interaction model.

§ System

The product combines a React client with typed service boundaries, durable conversation state, streamed model activity, provider-specific adapters, authentication, billing-aware entitlements and observability.

• Streaming keeps reasoning, tool work and final output legible.
• Provider adapters keep each model family's distinctive capabilities behind a common product model.
• Guest and signed-in paths use different model and usage boundaries.
• The same product is exercised across desktop and mobile.

§ Operating it

I own the product loop from interaction design and model routing through deployment, traces, incident response, and the unglamorous fixes that keep a multi-provider system coherent as APIs change.

§ Product

Each tenant gets a persistent agent with its own context, tools, schedules and delivery identity. The useful work is broad: research, images, documents, reminders, lightweight software and ongoing conversations. The interface remains a message thread.

§ Native transport

In 2026 I removed the former webhook architecture and moved the platform to OpenClaw's native imsg path. A broker resolves each inbound message to at most one tenant, isolates socket and attachment namespaces, and rejects ambiguous ownership.

• Compare-and-swap route changes stop a stale owner from winning.
• Per-tenant journals give bounded, at-least-once replay after an interruption.
• Health checks walk the real delivery path instead of trusting process state.
• Backups capture one coherent transport generation.

§ Evidence without surveillance

The plate is the public product page. Private conversations, phone numbers, account identifiers, the people behind the agents and the sites they build are not portfolio material.

§ Lifecycle

Provisioning creates an isolated tenant, applies a bounded tool and model policy, establishes delivery routes, and records enough state for the service to be inspected and repaired later.

§ Safety model

The runtime combines container and process isolation with explicit filesystem and credential scopes. Transport ownership fails closed, and health checks exercise the user-facing path rather than reporting a green process alone.

§ Operations

Deployment, logs, traces, usage controls, backups and repair workflows are part of the product. This is where most of the hard-won agent expertise lives: making stateful, tool-using software recoverable.

§ Why build it

The iPad is still missing the editor development needs. Tide came out of a research pass over about twenty-five iOS editors, VS Code's internals and iOS platform limits, and starts from the workbench people already know, rebuilt natively, before adding remote machines and agents.

§ What works today

SwiftUI drives the workbench and UIKit with TextKit drives the editor. The first version already carries a surprising share of a desktop editor.

• Incremental syntax highlighting for about thirty languages, using VS Code's line-state tokenization model.
• Quick Open with VS Code's fuzzy scorer and prefixes, project-wide search and replace, and an outline.
• Ten color themes, VS Code theme import, and a settings editor keyed to VS Code's setting names.
• A terminal panel and a pure-Swift reader for .git folders: branch, status and diffs against HEAD.

§ Next

The roadmap adds SSH hosts and remote workspaces, language servers over SSH, and AI assistance whose edits are reviewed hunk by hunk. It is public and MIT licensed, and labelled early on purpose.

§ Why build it

Terminal airspace is usually taught as a stack of flat diagrams. The rules make more sense when the learner can move a probe through the actual volume and see what changes with position and altitude.

§ Interaction model

The lab combines orbit and profile cameras, altitude slicing, an aircraft probe, class comparisons and short knowledge checks in a focused Vite application.

• Direct manipulation instead of a passive animation.
• Profile and plan views connect vertical and horizontal limits.
• Progressive controls keep it useful on smaller screens.
• Minimal dependencies keep the experiment easy to inspect.

§ What it is

A storefront with its own operations desk: products are imported, priced by margin rules, given generated ad copy and photography, and sold through Stripe checkout.

§ What it is

Discover, identify and share the natural world: a web app, an API and an iOS app over the same record. Signup is open on the web and on TestFlight.

§ What it is

A macroeconomics dashboard that charts FRED series and runs scenarios over them. It is presented as a product experiment rather than a verified data authority.

§ What it is

A Christian AI initiative, published as a zero-build static site on Cloudflare Pages.

§ What it is

A training log and coach that plans around the equipment your gym has, logs the work, and gives feedback on form from video. Web and iOS; the iOS app is in internal TestFlight.

§ What it is

A live music studio on Lyria RealTime: you steer the music while it plays, and the sketchbook keeps what you made so you can come back and edit it.

§ What it is

A companion for questions of faith that answers from Scripture, Tradition and the Magisterium, and checks every citation against its source before showing it.

§ What it is

An aviation platform for student pilots: an interactive E6B flight computer, a ground school you fly through, flight tracks, live ATC and weather decoded into plain language.

§ Workflow

A prompt can target one provider or fan out to a configurable council. Responses stream into stable panes, which makes differences in reasoning, citations, tone and failure behavior easy to inspect.

§ Current implementation

Rust, with a Ratatui interface, provider-aware routing, retries, web search, file and stdin context, and both interactive and composable command-line modes. ai-keys, its public sibling, manages the keys.

§ Design principle

The tool does not pretend consensus is truth. Its job is to expose independent outputs cleanly enough for a person, or a later synthesis step, to compare the evidence.

§ Problem

Multi-provider development scatters key creation, status checks, revocation and machine setup across unrelated dashboards and local configuration formats.

§ Prototype

ai-keys explores one CLI for provider-aware lifecycle commands and controlled distribution to machines. It is public and MIT licensed, and presented as a prototype rather than a finished distribution.

§ Interaction

A local wake-word model opens a full-duplex conversation. The implementation can route among realtime voice providers while keeping the same physical interaction.

§ What it tests

Voice exposes latency, turn detection, cancellation, audio-device recovery and cost in ways a text demo hides. The device is a compact test bed for those limits.

§ What it is

An original music video about AI culture, built with Blender and a simulation pipeline. The video link is still to come.

§ Constraint

Important seismic-analysis methods often live behind brittle legacy executables and hand-managed file flows. Replacing validated numerical work wholesale would add risk; wrapping it carelessly would keep the friction.

§ Approach

Excito puts a portable Python command line, a GUI and plots around a modernized RSPMatch (N. Abrahamson, 1993) and connects the results to a modern site-response workflow. The boundary keeps provenance visible while making batch runs repeatable.

• Explicit input and output contracts around the legacy routine.
• Scripted studies in place of manual desktop steps.
• Plots for checking each match by eye.

§ Merged work

Seventy-two pull requests landed in ooxml, the logic behind its Office viewers, and ten across OpenClaw's projects: eight in Clawpatch and one each in OpenClaw and imsg. Five more are open on OpenClaw.

§ Why these changes

Agent tools spend much of their life at external boundaries: model providers return partial data, network calls fail transiently, schemas evolve, and channels carry ambiguous context. These contributions make those edges explicit rather than hiding them behind a generic failure.

§ Live evidence

The evidence link opens GitHub's own search for merged pull requests by this author outside his own repositories, so the current record is inspectable at the source rather than asserted here.