TL;DR
- Pipper is an interface to control your agents. It’s not a harness: you bring your own subscriptions from providers like Claude Code, Codex, Cursor, and more.
- This article explains the thinking behind it: the assumptions Pipper was built on, how its adaptive interface splits into an orchestration layer and a workspace management layer, how the interface itself is laid out, and where I think these tools are heading.
Assumptions
Pipper was built around a few core assumptions to improve the developer experience:
- Tab bars and sidebars
- A tagging system to invoke a model
- By default, all agents run in YOLO mode
- The interface must adapt to your work
Adaptive Interface
We need to rewire our thinking when using these interfaces. They’re completely different from the chat interfaces of 2024–2025.
We need to reimagine the interface itself: instead of requiring you to adjust to the tool, the tool must adapt to your workflow. In Pipper, the interface molds to your workspace and guides you to the next best action. It is divided into two parts:
- Orchestration layer
- Workspace management
Orchestration in Pipper is different from what you’ve seen in other harnesses. We don’t
give models tools like read_file or edit_file; the models
running in Pipper get those from their own harnesses.
Pipper exposes every action a user can take as a tool via MCP, such as creating threads, sub-threads, and workspaces. It also exposes connectors: the agent harnesses (Claude Code, Codex, Cursor) and external tools (GitHub, Slack, Linear) you’ve connected to Pipper.
So whenever you’re stuck, just ask your agent. It knows more than you’d expect.
Based on the current state of your workspace, interface elements shift color and present the next relevant action.
Interface
The interface plays a big role in your productivity and the number of PRs you ship. Just switching from Claude Code or Codex to Conductor can help you ship more PRs.
Conductor is designed to be worktree-first, encouraging you to spin up isolated worktrees to run agents in parallel, whereas tools like Claude Code and Codex nudge you toward creating more threads.
Pipper’s adaptive interface is a hybrid that sits between the two.
Tab bar vs. sidebar: personally, I’m a tab bar advocate, but Pipper uses both, pairing two sidebars with a primary tab bar. The left sidebar holds only projects, while the tab bar houses threads, terminals, and previews. The right sidebar is a quick-glance panel showing the agent’s to-do list, CI checks, file tree, sub-threads, and workspace details.
The agent interface in Pipper is like a free canvas where you can tag @model inline to switch models and @terminal to run commands.
The interface also supports inline HTML rendering, so you get interactive reports instead of long walls of text.
Future
Here are my two predictions for how interfaces and these developer tools will evolve:
- By March 2027, the tools we build with will bend to how each person works rather than forcing everyone into the same layout. Every developer will run their own modified copy, and updates will arrive as small changelogs that an agent merges cleanly into their personal version.
- Siri will be the right way to control your agents when you’re away from your screen.