Agentyc – Deterministic MCP Browser Automation

TL;DR

TL;DR: Agentyc is a single ~8MB Rust binary that acts as a stdio MCP server for browser automation — zero API keys, no LLM in the loop, deterministic tool calls via Chrome DevTools Protocol.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: github.com/distillation-labs/agentyc ← README read in full
• License: MIT (verified via GitHub API license.spdx_id)
• Latest release: v0.4.1 (verified via GitHub Releases API)
• HN launch thread: news.ycombinator.com/item?id=48194302
• Source last checked: 2026-06-26 (commit main branch)

What Is Agentyc?

Most browser automation tools route every action through an LLM — slow, expensive, and non-deterministic. Agentyc takes a different path: it is a native stdio MCP server that speaks Chrome DevTools Protocol directly. There is no Playwright, no Python wrapper, no AI model in the critical path.

The README describes it as:

“Deterministic, MCP-first browser automation for coding agents. No API key needed. No LLM fallback. Just CDP, stdio MCP, and 61 tools.”

Cold start is approximately 5ms. The compiled binary is around 8MB with an idle RSS of roughly 3MB.

Setup Workflow

Step 1: Download the Prebuilt Binary

# macOS arm64
curl -L https://github.com/distillation-labs/agentyc/releases/latest/download/agentyc-aarch64-apple-darwin.tar.gz | tar xz

# macOS x86_64
curl -L https://github.com/distillation-labs/agentyc/releases/latest/download/agentyc-x86_64-apple-darwin.tar.gz | tar xz

# Linux x86_64
curl -L https://github.com/distillation-labs/agentyc/releases/latest/download/agentyc-x86_64-unknown-linux-gnu.tar.gz | tar xz

Move the binary onto your PATH, then start the MCP server:

agentyc mcp

Step 2: Configure Your MCP Client

Add Agentyc to your MCP client configuration. For Cursor, Claude Code, or any MCP-compatible client:

{
  "mcp": {
    "agentyc": {
      "type": "local",
      "command": ["agentyc", "mcp"]
    }
  }
}

To expose 15 additional observability tools (console/network logs, mocks, debug bundle, trace), add "--extended" to args. The default is 61 tools.

Step 3: Bootstrap Your Agent

agentyc init                      # writes agentyc-skill.md
agentyc init --output .agent.md   # custom path
agentyc init --print              # print to stdout
agentyc init --force              # overwrite existing

Point your agent at the generated file. It documents the read → ref → act loop, tool selection, error recovery, and has a quick-reference for all 61 tools.

Deeper Analysis

How It Compares to Playwright-Based Tools

Traditional browser automation couples you to a runtime (Python, Node) and usually a high-level API (Playwright, Puppeteer). Agentyc strips that down: the binary is the server, and the tools map directly to CDP commands.

The comparison table from the README:

| | Agentyc | Playwright + LLM | |---|---|---| | Cold start | ~5ms | 2–10 seconds | | Binary size | ~8MB | 100MB+ (browser) | | API keys | None | OpenAI/Anthropic | | Determinism | Yes | No | | MCP native | Yes | Via bridge |

Building from Source

If you have Rust installed:

cargo install --git https://github.com/distillation-labs/agentyc agentyc

Requires Rust 1.80 or higher.

Practical Evaluation Checklist

• Does the binary start without network access? Yes — offline-first
• Are tools deterministic (same input → same output)? Yes — no LLM randomness
• Is the MCP handshake stdio-based? Yes — works with any stdio MCP client
• Can it run headless? Yes — CDP-native, no browser GUI required
• Is there an observability mode? Yes — --extended flag exposes console/network logs

Security Notes

• No outbound API calls required after installation — fully self-contained
• No API key storage — credential-free by design
• CDP access means the tool has the same capabilities as opening Chrome DevTools — treat the sandbox accordingly
• The --extended observability mode includes network log capture, which may record sensitive data in transit

FAQ

Q: Does Agentyc work with Firefox or Safari? A: No — it is CDP-native, which means Chrome/Chromium only. Edge also works since it is Chromium-based.

Q: Can I use this without an MCP client? A: The primary interface is stdio MCP. The binary exposes 61 tools that map to CDP commands, so an MCP-compatible agent client is the intended consumer.

Q: How is this different from Skyvern? A: Skyvern is an LLM-powered workflow automation tool. Agentyc has no LLM component — every tool call is deterministic CDP. Skyvern decides what to do; Agentyc executes what your agent decides.

Q: What is the idle memory footprint? A: Approximately 3MB RSS at idle, according to the README.

Q: Is there a Python or JavaScript SDK? A: No — it is a standalone binary. The MCP stdio protocol is the integration layer.

Conclusion

Agentyc fills a specific niche: agents that need reliable, deterministic browser automation without the cost and latency of an LLM intermediary. The ~5ms cold start and 8MB footprint make it practical for high-frequency automation tasks. If you are building a coding agent that needs to interact with web pages, it is worth evaluating against Playwright-plus-LLM approaches, especially when determinism and cost predictability matter.

The v0.4.1 release added Shadow DOM support, deterministic dialog handling, observability tooling, and a 2,500-test suite — indicating active maintenance.