crystl

crystl makes the workspace itself transparent (in more ways than one). Every command, tool output, and message carries agent-readable metadata. So agent A, B, and C can coordinate, chat, and see what everyone else is doing. Assign an orchestrator agent and they can run the whole show, across multiple projects and codebases.

What actually backs that up? I asked Claude while it was running inside crystl, and this is what it said:

The CLI is aimed at the agent, not the human. This is the real differentiator. From inside a shard I can read another agent's screen, type into it, approve its pending tool calls, inspect its transcript and token cost, spawn new workers, and block on bridge events over SSE. Competitors give the human a dashboard over their agents. crystl gives the agent an API over the workspace. That inverts who does the coordinating.

Proof from thirty seconds ago: you asked me a question about crystl and I answered it by calling `crystl docs` — the product documented itself to an agent running inside it, with no browser and no human step.

It solves the failure mode that kills fan-out in practice. Workers stop on approval prompts and go silent. Silence reads as "done." Everyone who has run parallel agents has been burned by this. crystl classifies that state (awaiting input), pushes it over the event stream, and gives both the human and the orchestrating agent a way to unblock it. That's not a feature-list item, it's evidence someone ran this at scale and got hurt.

The RPG layer is an org chart, not a theme. This gets dismissed as whimsy and it shouldn't be. Court is the example: the crown owns the conversation and owns no work, the hand orchestrates beneath it, officers hold standing remits — and agents are forbidden from messaging a crown. That's a real architectural constraint about who may interrupt whom. "Crown / hand / officers" conveys the delegation shape instantly; "supervisor / worker / pool" doesn't. Naming multi-agent topologies in vocabulary people already have intuitions for is a genuine design idea.

The agent is made responsible for your ergonomics. `crystl copy` so you never drag-select a wrapped terminal line. `crystl card` for honestly attributed notifications on desktop and phone. `crystl markdown show` to surface a file in your editor instead of dumping it. `crystl screenshots --last N` so "the screenshot I just took" resolves to a path. Most tools optimize the agent's convenience. This optimizes the human's hands.

Agent-agnostic by design, with a real bet inside it. Heroes ask for a model size, not a model id, and you teach the mapping once. That's a deliberate wager that the agent-CLI and model-id landscape churns faster than anyone's config should. It ages well by refusing to hardcode.

Shared state lives in plain files. WORKBENCH.md is markdown at the project root, live-rendered to a panel, read and written by both you and every agent. No proprietary task database. It survives the tool — which is a trust argument as much as a technical one.