lodeflow

The workspace for directing agents

The way you build changed.
Your tools didn't.

You direct agents now; you don't hand-edit files. LodeFlow is the calm workspace built for it: host the real CLIs, run a fleet across every project, stay in your flow.

See how it worksRepo
claude · acme-api▲ needs you
> make checkout_test deterministic — it
  flakes ~1 in 7 on CI

⏺ Reproducing first, then isolating the race.

⏺ Bash(npm test -- checkout_test --runs 20)
  ⎿ 20 runs · 3 failures · seed set in parallel

Do you want to proceed?
❯ 1. Yes
  2. Yes, and don't ask again
  3. No, tell Claude what to do differently
Several agents, many projects, one glance. The real loop.Representative; a live capture replaces this at launch.

You're directing agents across a dozen projects from a sprawl of terminal windows. Juggling which login still has quota. Losing your place on every restart. Dropping into an IDE built for hand-editing code you don't write anymore. It works. It's also a disaster.

It doesn't have to be. The way we build changed; the tools didn't. LodeFlow is built the other way around, for directing agents, not typing code.

Terminal-native. The agent's own CLI is the best interface anyone will build for it. So LodeFlow doesn't reimplement Claude Code or Codex. It hosts the real binaries in real terminals and never gets between you and the agent. No imitation chat box; the thing itself.

Multi-agent-native. One agent is a demo. The real day is several at once, across every project you touch: the work codebase and the side project, side by side, each readable at a glance. Built around a fleet from the first commit, not a single-document editor with a chat panel bolted on.

The whole product

Everything directing agents needs. None of the IDE you don't.

That's one piece. The rest of the loop is here too: assembled and working as one, not a host plus four promises.

  1. It just works

    The agent's CLI is the best interface for it. So host the real one, don't rebuild it.

  2. Direct the fleet

    One agent is a demo. Directing ten across your projects is the job.

  3. Sort your subscriptions

    Multiple accounts, live quota, routing to the right one. Your own paid plans, ToS-safe.

  4. Review the work

    You stopped typing code. Reviewing the decisions is the new work.

  5. A project that learns

    Most tools forget between sessions. This one compounds: every session starts smarter.

Markdown is where you live

You don't write code.
You write markdown.

Plans, notes, decisions, and the project's own memory: all markdown, with a real editor and scroll-locked live preview. It's the surface vibe coding actually happens on, and we build for it like it matters.

Is

A native desktop workspace for vibe coders directing coding agents. It hosts the real agent CLIs (Claude Code, Codex) faithfully in real terminals, never rebuilding their UI. It runs a fleet of them across all your projects, sorts your subscriptions so you're never throttled, lets you review what they did, and learns your project so every session starts smarter than the last.

Isn't

Not an IDE. It doesn't instrument around code: no LSP, refactoring, debugger, autocomplete, or minimap. Change a line if you want; LodeFlow just doesn't build the machinery for it. Directing agents is the job, not typing code.

Who's building this

Hi, I'm Nathan. I'm building LodeFlow solo, and I direct agents all day. This is the calm workspace I wanted and couldn't find.

Find me on GitHub

The name

A lode and a workload sound identical and mean the opposite: a rich vein of ore worth mining, not a pile to grind through. LodeFlow is for working the vein, and staying in flow.

And yes, that's why everything here is the color of ore.

The thread

The devlog is the long-form answer.

Architecture decisions, how agents actually retrieve memory, the dead ends. Judgment narrated as it's made. The real interview is async.

Read the devlog →