Try CLI — interactive terminal demos

Try CLI is an AgentClash platform primitive: interactive, disposable terminal demos for CLI and TUI tools — including AI coding agents.

Let users try your CLI before they install it — not a cloud IDE, not “run any repo.”

Where it lives

The browser UI ships on Vercel with the AgentClash web app. PTY sessions run on a long-lived Bun service backed by E2B sandboxes (WebSockets cannot run on Vercel serverless). The service ships deploy configs for both Railway (railway.toml) and Fly.io (fly.toml).

Demos

AI coding agents — Claude Code, Codex CLI, OpenCode, Grok CLI, Kimi K2, Qwen3 Coder. Developer tools — bun, uv, ruff, biome, ripgrep.

Every tool is pre-installed in a shared E2B template (agentclash-trycli), so sandboxes boot ready in ~1–2s instead of installing on each visit. Rebuild the template after changing tools:

1cd services/try-cli
2E2B_API_KEY=... bun run scripts/build-template.ts

The build runs in E2B's cloud (Build System 2.0 — no local Docker). The demo's template: field references the alias.

Auth: free demo, then bring-your-own

Anonymous free demo. For agents that can run on AgentClash credentials, visitors can try them for a few minutes with no login and no key. The CLI is wired to a metered server-side gateway that holds AgentClash's provider keys, mints a short-lived spend-capped token per session, and never lets the real key into the sandbox. A durable Redis-backed daily ceiling is the hard backstop. Grok CLI is the exception — it is key-only, so you bring your own xAI API key from the first session.

Then sign in. When the demo ends, signing in with AgentClash gives a longer session where you authenticate the CLI with your own account/key (running on your quota, not ours). We never inject our keys into the sandbox — that would violate provider terms and expose the key to anyone in the shell. Each demo's auth: block shows how to sign in with your own credentials:

• Claude Code — claude → /login (paste-code) or ANTHROPIC_API_KEY
• Codex — codex login --device-auth or an OPENAI_API_KEY
• OpenCode — opencode auth login (paste a provider key)
• Grok — export XAI_API_KEY=…

Credentials live only in that sandbox and are wiped when it expires.

Maintainer workflow

Add .trycli.yml to your repo:

1name: "my-cool-cli"
2install:
3  - "curl -fsSL https://example.com/install.sh | bash"
4welcome: |
5  my-cool-cli is ready. Try: mycool --help
6commands:
7  - label: "Show help"
8    run: "mycool --help"
9sessionMinutes: 10

Publish a badge:

1npx @agentclash/try-cli init
2npx @agentclash/try-cli publish

[![Try on AgentClash](https://www.agentclash.dev/api/try/badge/mycool.svg)](https://www.agentclash.dev/try/mycool)

Challenge packs & terminal tool kind

A challenge pack's tool_policy.allowed_tool_kinds is a list of tool kinds — the valid values are browser, build, data, file, network, and terminal. The terminal kind marks packs that expect interactive terminal / Try CLI surfaces. Shell execution is a separate primitive, not a tool kind: it is gated by the tool_policy.allow_shell flag, so putting exec in allowed_tool_kinds fails validation.

Use shell (allow_shell: true) inside the sandbox to validate CLI behavior in eval runs; use Try CLI for human-facing README demos.

Architecture

web (Next.js / Vercel)        →  /try, /api/try/* proxy
services/try-cli (Bun service)→  E2B PTY ↔ WebSocket
services/try-cli/scripts      →  build-template.ts (E2B template build)
try-cli/packages/core         →  .trycli.yml schema + badges
try-cli/demos                 →  Curated demos (AI agents + dev tools)

Session limits (MVP)

• 10 minute sessions (configurable per demo)
• Max 3 concurrent sessions per IP (freed when a session ends)
• Sandboxes destroyed on expiry — credentials wiped with them
• Anonymous; AI agent demos use the visitor's own provider credentials (BYO)

See also

• Sandbox & E2B
• Tools, primitives & policy
• Repo: try-cli/ and services/try-cli/
• Datasets overview
• Multi-turn packs
• Security evaluation