_{Follow me on X @ybouane — I'm building in public.}

Make Claude Scriptable Again

dash-p

If Claude can interact with apps through their interfaces, developers should be able to interact with Claude Code through theirs.

Claude can click buttons, type into apps, read screens, and operate software through a GUI.
So why shouldn’t developers be able to automate their own Claude Code session?

dash-p is a CLI and TypeScript library that makes Claude Code scriptable by driving the real Claude TUI process you already use.

It does not bypass Claude Code.
It does not fake network requests.
It does not bypass authentication.

It simply launches the official claude command, injects input programmatically, reads output programmatically, and exposes a clean developer interface on top.

The same Claude Code.
The same local session.
The same authentication flow.
Just composable.
Not a replacement for Claude Code.

A bridge that makes it scriptable.

Use it in one line

CLI — replace claude -p:

- claude -p "summarize this repo"
+ dash-p  "summarize this repo"

SDK — replace the import:

- import { query } from "@anthropic-ai/claude-agent-sdk";
+ import { query } from "@ybouane/dash-p";

That’s the whole idea. Same prompt in, same shape of answer out — but produced by driving the interactive TUI instead of the -p headless flag.

Getting started

Prerequisite: the official claude CLI, installed and logged in — run claude once interactively to sign in. dash-p drives that binary; it does not bundle or replace it. (Also: Node ≥ 20. node-pty installs a prebuilt binary on common platforms; if yours needs compiling you’ll want a C++ toolchain — macOS xcode-select --install, Debian/Ubuntu apt install build-essential python3.)

CLI — install globally:

npm install -g @ybouane/dash-p

dash-p "what color is the sky"

That puts a dash-p command on your PATH. Verify with which dash-p. If claude isn’t on your PATH, pass --claude-path /abs/path/to/claude.

SDK — add to a project:

npm install @ybouane/dash-p

import { query } from "@ybouane/dash-p";

No install:

npx @ybouane/dash-p "what color is the sky"

CLI

dash-p "explain this repo in two sentences"
dash-p -m sonnet "summarise the CHANGELOG"
dash-p -o json "what is 6 * 7? reply with just the number"
echo "a long prompt from a pipe" | dash-p
dash-p -o stream-json "name two primary colors"     # JSONL message stream

Output formats mirror claude -p: text (default), json, stream-json. Run dash-p --help for the full flag list (model, permission mode, tools, working dir, terminal size, and more).

SDK

The query() shape matches @anthropic-ai/claude-agent-sdk, so code ports over:

import { query } from "@ybouane/dash-p";

for await (const msg of query({
  prompt: "In one sentence, what is a pseudo-terminal?",
  options: { model: "sonnet", includePartialMessages: true },
})) {
  if (msg.type === "stream_event" && msg.event.type === "content_block_delta")
    process.stdout.write(msg.event.delta.text);
  if (msg.type === "result") console.log("\n[done]", msg.result);
}

You get system / user / assistant / stream_event / result messages, with structured tool_use + tool_result blocks. See docs/SDK-PARITY.md for exactly what’s supported.

How it works

 your CLI / query()  →  dash-p engine  →  spawns the real `claude` in a PTY

1. Launch — spawns the official claude interactive TUI inside a pseudo-terminal (so it renders its true UI, exactly as for a human).
2. Inject — pastes your prompt and keystrokes into the terminal.
3. Read — feeds the terminal output through a headless terminal emulator (the same engine VS Code uses) to reconstruct the screen.
4. Extract — recognizes the conversation structure (assistant text, tool calls, results) and returns it as clean text or structured SDK messages.

All the Claude-specific knowledge (what the prompt box looks like, how a tool call renders, etc.) lives as data in versioned profiles/, so a TUI update is a profile edit, not an engine rewrite. There’s even a self-recalibration skill for new Claude releases.

For the full architecture, see docs/DESIGN.md.

Good to know

dash-p reads the rendered screen, so it’s inherently tied to how the TUI looks (it’s calibrated against Claude Code 2.1.x). A few things are fidelity ceilings of screen-scraping — e.g. markdown syntax like code fences is rendered-then-lossy.

For byte-exact output when you need it, dash-p can read the session transcript Claude Code already writes to disk:

dash-p --enrich-from-session "..."   # exact text + token usage
dash-p --verify-session       "..."   # warn if the scrape diverges from ground truth

This stays read-only — dash-p still drives only through the TUI.

This is a research experiment, not an official product. Use your own account and respect Claude Code’s terms.