// Tools & Experiments
Local-first tools for agent-heavy work.
These are the small systems I build around my own workflow: Codex remote control, session recall, mailbox evidence, tool-feedback loops, and desktop automation. They sit behind the client work, but they show how I think about reliability, permissions, and operational context.
agent surface codex, telegram, mailbox, browser
state sqlite, jsonl, markdown, local artifacts
rules narrow permissions, receipts, approval gates
output tools that make real work easier to repeat
// Featured Tools
Useful enough to explain publicly.
The point is not a long list of side projects. These are the tools that best show the operating system I am building around agentic work.
Remote Codex control
codex-telegram-bridge
A local bridge that lets me mark myself away, receive Codex updates in Telegram, reply back into the right thread, inspect active work, and approve actions without keeping the laptop in front of me.
It turns Codex Desktop from a single-machine workflow into something I can safely supervise while away.
Session recall
codex-recall
A local SQLite FTS index over Codex session archives, built so agents can search prior work, inspect receipts, and reuse context without treating raw JSONL logs like a durable database.
It gives long-running agent workflows memory with evidence instead of vague recollection.
Mailbox evidence
mailbridge
A conservative JSON CLI over local Gmail and Fastmail access for search, read, thread lookup, attachment download, evidence pulls, and draft-only composition.
It gives agents useful mailbox context while keeping the boundary narrow: no send, archive, delete, or label mutations.
Tool improvement loops
tool-feedback
A local runtime for submitting tool-friction cases, deduping them, routing triage or patch jobs, and sending approval/status updates through Telegram.
It makes tool improvement a tracked workflow instead of scattered notes after something breaks.
macOS automation
cli-tools
A TypeScript monorepo of local CLIs for macOS apps, Granola notes, browser tabs, and social workflows that agents can call during real work.
It keeps useful desktop context reachable through explicit commands instead of fragile one-off scripts.
// Design Principles
The same judgment shows up in client systems.
The useful state stays on the machine whenever possible: transcripts, mailbox pulls, tool feedback, and workflow databases.
Tools should expose the smallest useful surface, especially around email, browser sessions, client context, and agent actions.
Agent workflows get better when they can point to source files, command output, run artifacts, and explicit state transitions.