hobkin
Get early access

Your coding agent waits for you.
hobkin doesn't.

hobkin runs on top of the agents you already use. Tell it what you keep doing, like the email you forgot to answer, the PR that needs a rebase or the plan you write every Monday, and it takes over on a schedule.

Get early access See how it fits

Runs on Pi. In Herdr, work mode also runs on Claude Code or Codex.

What it is

Not a new agent. A layer on the ones you have.

You already picked your harness, tuned your setup and pay for your plan. hobkin doesn't ask you to switch. It drives the agents you already use, adds its own memory, and works on its own schedule instead of waiting for your next prompt.

Keeps your setup

No new agent to learn, no config to move. hobkin works through your Pi tools, extensions and models.

Works across all of them

Chat and routines run on Pi. Work mode runs on Pi, Claude Code or Codex, and you can switch whenever you like. hobkin learns from the runs on each one.

Gets better every round

Each run leaves notes behind, so the next one starts further ahead.

Routines

The things you keep forgetting, handled.

Tell hobkin what to watch and when: unanswered emails, PRs that need a rebase, the plan you write every Monday. Each check runs on the model you pick and remembers what it found last time, so it only speaks up when something changed.

The Routines tab, with example checks. hobkin runs while it's open, and on a Mac it keeps the machine awake while the screen sleeps.

Quiet unless it matters

Findings worth your time show up in Chat. Runs with nothing new stay in Reports.

Picks up where it left off

Every check keeps notes on what's open and what it already told you, so you don't get the same news twice.

Stops before it breaks things

After two failed runs in a row, hobkin pauses the check, finds the cause and brings you a fix. Nothing changes until you approve it.

Work mode

Twenty side projects. All of them moving.

Running more projects than you have evenings? Type /work, pick the ones you care about and set their priority. Agents keep each one moving in small, reviewable changes. The vibe-coded prototype gets its tests, the stale repo gets its dependencies bumped, and the bug you noted last month gets fixed.

  • Security
  • Tests
  • Features
  • Bugs

The Work mode tab, with example projects. In Herdr you can follow each agent in its own pane, on Pi, Claude Code or Codex.

Every task earns its place

An agent can't file a vague idea. It has to show the evidence, say who benefits, name the first change and explain how it will prove the work is done.

You stay the reviewer

Every task starts with a failing test or check and ends on its own branch, ready for review. Agents never commit to your main branch, and nothing is pushed unless you ask.

Your plan stays yours

Work mode has a daily token limit. On Pi it also leaves 30% of each usage window for you and paces itself across the week.

Learning

It learns from you. You learn from it.

Turn learning on and hobkin watches how work turns out: which reports you open, which proposals you take, which PRs you merge. It turns that into short lessons every agent follows. Give it your past sessions and it can show you the patterns you'd miss.

Tell it once

Say how you like things done and every agent follows it, in chat, routines and work mode. Memory is stored on your machine.

Spots your repeats

Pick past Pi or Claude Code sessions and hobkin lists the requests you keep repeating, with a routine for each one you could hand off. Nothing starts until you say so.

Learning costs nothing

Lessons are worked out without a model, so they use no tokens.

What the lessons look like
learned: advice only, grants nothing
-site-audit blocked x5: do the rest, name what stopped it
-news-digest unread x5: lead with the change, be short
+price-watch used x5: keep this approach

Early access

Be among the first to try hobkin.

hobkin is in private release for now. Leave your email and I'll write to you once, when it's ready.