Reverse API Engineer MCP Server

<div align="center"> <img src="https://raw.githubusercontent.com/kalil0321/reverse-api-engineer/main/assets/reverse-api-banner.png" alt="Reverse API Engineer Banner"> <br><br> <a href="https://pypi.org/project/reverse-api-engineer/"><img src="https://img.shields.io/pypi/v/reverse-api-engineer?style=flat&color=e50d75&labelColor=1f1f1f" alt="PyPI"></a> <a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.11+-e50d75?style=flat&labelColor=1f1f1f" alt="Python"></a> <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/license-MIT-e50d75?style=flat&labelColor=1f1f1f" alt="License"></a> <br>

</div>

Reverse API Engineer

Turn websites into APIs. Browse (or let an agent browse), and get a clean, typed client for the endpoints the site actually uses.

<p align="center"> <img src="https://raw.githubusercontent.com/kalil0321/reverse-api-engineer/main/assets/rae-agent-steam.gif" alt="Agent mode turning Steam's charts page into an API client"> <br> <em>Agent mode</em> </p>

---

How it works

1. You give it a website and a goal ("fetch all Apple jobs").
2. A browser visits the site, either driven by you or by an AI agent.
3. Network traffic is captured to a HAR file.
4. Your configured model reads the traffic and writes a working API client in Python, JavaScript, TypeScript, Go, Java, C#, PHP, Ruby, C, or PowerShell.

No more manually opening DevTools, copying cURL commands, and gluing together a client.

Install

uv tool install reverse-api-engineer   # or: pip install reverse-api-engineer

The secured cryptography dependency no longer supports Intel macOS or 32-bit Windows. Use Linux (including on Intel hardware), Apple Silicon macOS, or 64-bit Windows. Installing on the retired platforms may attempt an unsupported source build requiring Rust and OpenSSL development tools; this is not a supported installation path. Do not downgrade cryptography to restore old wheels. See the upstream platform changes.

Agent mode (the default) captures through npx-launched browser tooling — browser MCP servers (Playwright or Chrome DevTools) or the Vercel agent-browser CLI — and needs no extra Python packages. Manual mode drives a local Playwright browser, which ships in the optional [manual] extra:

uv tool install "reverse-api-engineer[manual]"   # or: pip install "reverse-api-engineer[manual]"
playwright install chromium

Use it from your coding agent

Install the skill so Claude Code, Codex, Cursor, OpenCode and other agents know how to drive the CLI:

npx skills add kalil0321/reverse-api-engineer

Then ask your agent for what you need, for example "get me a typed client for the job search on this careers page". The skill lives in skills/reverse-api-engineer.

Quick start

reverse-api-engineer
> fetch all apple jobs from their careers page

# Browser opens. Navigate, interact, close when done.
# → ./scripts/apple_jobs_api/  (api_client.py, README.md, example_usage.py)

Cycle modes with Shift+Tab:

Mode	What it does
manual	You drive the browser; AI generates the client from captured traffic.
agent	An AI agent drives capture autonomously (Playwright or Chrome MCP, or Vercel agent-browser CLI).
engineer	Re-run generation on a previous capture (engineer <run_id>). In the session it works on your latest run with a HAR, named above the input; /engineer picks another. Needs a HAR, so it works on auto, agent-browser and manual runs, not chrome-mcp runs.

Agent mode providers:

• chrome-mcp (default for new configs; an existing config.json keeps its provider): drives your real Chrome so you keep existing sessions/cookies. Needs Node.js 20.19+; headed runs also need Chrome 146+ open with auto-connect enabled at chrome://inspect/#remote-debugging, and someone at the machine to accept Chrome's connection prompt when the run starts. For headed runs, rae checks before launching that Chrome's remote debugging port is reachable (not the Chrome or Node version) and stops with setup steps if it isn't; agent --dry-run shows the same check. Headless runs skip it. --headless starts a separate headless Chrome without your sessions. No HAR file is saved, so engineer <run_id> cannot regenerate these runs; run agent again instead.
• auto: Playwright MCP, records a HAR and needs no Chrome setup.
• agent-browser: Vercel agent-browser CLI (not a Reverse API Engineer browser MCP server). At session start RAE uses whatever agent-browser is already on PATH, otherwise runs npm install -g <pin> (same pin as config / RAE_AGENT_BROWSER_PACKAGE), prints a yellow notice, validates with --help, and only then falls back to npx -y <pin> if npm cannot install. Prompts embed the resolved shell prefix alongside skills get core --full, skills list, HAR phases, cloud notes from agent_browser_notes. Tune with agent_browser_npx_package (optional), env RAE_AGENT_BROWSER_*. First Chromium fetch: agent-browser install (add --with-deps on trimmed Linux).

Optional sanity checks:

agent-browser doctor --offline --quick || true
agent-browser skills list >/dev/null

Configuration

Settings live in ~/.reverse-api/config.json and can be edited via /settings in the CLI:

{
  "agent_provider": "chrome-mcp",
  "agent_browser_npx_package": "agent-browser@0",
  "agent_browser_notes": "",
  "claude_code_model": "claude-sonnet-4-6",
  "claude_setting_sources": [],
  "ollama_auto_start": true,
  "ollama_base_url": "http://127.0.0.1:11434",
  "opencode_auto_start": true,
  "opencode_base_url": "http://127.0.0.1:4096",
  "opencode_model": "big-pickle",
  "hermes_max_turns": null,
  "hermes_model": null,
  "hermes_provider": null,
  "opencode_npx_package": "opencode-ai@latest",
  "opencode_provider": "opencode",
  "copilot_model": "gpt-5",
  "codex_model": null,
  "codex_effort": null,
  "codex_home": null,
  "cursor_model": "composer-2.5",
  "output_dir": null,
  "output_language": "python",
  "real_time_sync": true,
  "sdk": "claude"
}

• Models: Sonnet 5.5, Opus 5.5 (most capable), Opus 4.6 and Haiku 5.5 (fastest). Sonnet 4.6 stays the default until you choose another; it and Haiku 4.5 are still accepted in config and by --model. For OpenCode see models.dev.
• SDK: claude (default), opencode, codex (OpenAI Codex), cursor, copilot (GitHub Copilot), or hermes.
• Hermes: engineers a saved capture through the hermes CLI instead of an SDK. It needs Hermes Agent v2026.9.21 or later from the official installer (the PyPI package is older): curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash on macOS, Linux and WSL2, iex (irm https://hermes-agent.nousresearch.com/install.ps1) in PowerShell on Windows. It needs no API key in rae: the model, the provider and the authentication are your own Hermes configuration. hermes_model and hermes_provider override which model that configuration runs; leaving them null uses whatever Hermes already selects. hermes_max_turns overrides the turn cap from your Hermes config. Runs start with --ignore-rules, so an AGENTS.md or CLAUDE.md in the workspace cannot steer the run. Cost is estimated from the token counts the CLI reports, and stays at 0 for a model with no known price. It covers engineer only: while sdk is hermes the agent command stops before opening a browser, so capture with claude, opencode, copilot or cursor, then switch sdk back to hermes to engineer that run. agent --dry-run reports the same thing.
• Claude settings isolation: Claude sessions don't load your Claude Code settings files, CLAUDE.md, hooks, plugins, permission rules or MCP servers, so your personal setup doesn't steer the agent. Organization-managed policy still applies, and Claude Code still reads its own state such as ~/.claude.json and auto memory. If ~/.claude/settings.json provides your authentication (apiKeyHelper, or Bedrock, Vertex or proxy variables under env), set claude_setting_sources to the layers to load, e.g. ["user"] or ["user", "project", "local"]. Each layer you list is loaded in full, including its hooks, plugins, permission rules and MCP servers.
• OpenCode setup: with sdk: "opencode", RAE reuses an existing server or downloads/starts opencode-ai@latest through npx; a global OpenCode installation is not required. Fresh configurations default to the free opencode/big-pickle model. /settings shows a loading list, then offers a searchable opencode provider / model list populated from the server's connected, tool-capable catalog and marks free OpenCode options. Before creating a session, RAE validates the saved pair again and suggests currently available free models when configuration is invalid. Compatible older servers are reused with a version warning. Node.js 20+ is required for automatic startup. Password-protected servers use OPENCODE_SERVER_PASSWORD and optional OPENCODE_SERVER_USERNAME. Override startup with OPENCODE_BASE_URL, RAE_OPENCODE_PACKAGE, or RAE_OPENCODE_AUTO_START=0.
• Codex setup: install with uv tool install "reverse-api-engineer[codex]" (it brings its own Codex runtime) and sign in once with codex login or the Codex app. RAE runs Codex in its own home, ~/.reverse-api/codex, whose auth.json links to your Codex sign-in, so your Codex config.toml, MCP servers, hooks, skills and AGENTS.md, and the apps connected to your ChatGPT account, don't steer the agent and RAE's runs stay out of your Codex history. Set codex_home (for example "~/.codex") to run in a Codex home as it is, with all of its setup; that is also how to share a sign-in kept in the system keyring, or on Windows without permission to create symbolic links. codex_model: null uses your account's default model; /settings lists the models your account offers. codex_effort: null uses the model's default reasoning effort; /effort lists the levels the model offers. Like the other backends, the agent runs without Codex's sandbox or approval prompts.
• Ollama through OpenCode: choose provider ollama in /settings; RAE starts an installed Ollama daemon if needed, lists only installed models with tool calling and 64k+ context, and supplies OpenCode's provider config inline. Models are never downloaded silently. Override with RAE_OLLAMA_BASE_URL or RAE_OLLAMA_AUTO_START=0.
• Output language: python, javascript, typescript, go, java, csharp, php, ruby, c, or powershell. C needs a POSIX toolchain (cc, libcurl headers) — macOS/Linux, or WSL/MSYS2 on Windows. PowerShell needs pwsh 7+ (PowerShell Core, not Windows PowerShell 5.1). Its runner executes the generated Example.ps1 beside api_client.psm1; argument overrides are not supported. Edit the example or import the module to call its exported functions with your own arguments.

CLI

reverse-api-engineer opens a full-screen session; agent and engineer open it too when run in a terminal, and stay plain with --json or --no-interactive. Type / for the commands:

• /model, /effort, /mode: switch the model, how hard it thinks, and the mode (also Shift+Tab, ctrl+e).
• /engineer [run_id]: the run engineer mode works on, by default the latest with a HAR.
• /settings: SDK, agent provider, models, output and sync, changed in place in the session.
• /history: past runs with their details, to resume or re-engineer.
• /resume [run_id]: show a past run's conversation as it was and continue it where it stopped, on any SDK. Works for agent and engineer runs from this version on, which save their SDK session.
• /output (ctrl+o): everything from the session, unfolded. /open: the latest client.
• /help (or /commands): what rae does, its modes, commands and keys.
• /new, /clear, /exit (or /quit).

Scriptable subcommands (pipe to jq):

reverse-api-engineer agent --prompt "capture the public jobs api" \
  --url https://example.com/jobs --json | jq

reverse-api-engineer list --json
reverse-api-engineer show <run_id> --json
reverse-api-engineer run <run_id> --file api_client.py \
  --no-interactive --auto-install -- --org acme

Pass --no-interactive (and/or --json) to skip prompts. With --json, stdout is one JSON document and logs go to stderr.

agent --json schema

Field	Type	Notes
schema_version	int	Currently 1.
status	"ok" | "error"	Top-level result.
run_id	string | null	Use with show / engineer / run.
prompt	string
url	string | null
mode	string | null	"auto", "chrome-mcp", or "agent-browser".
har_path	string | null	Captured HAR.
script_path	string | null	Generated client; exists whenever status is "ok".
usage	object	{input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, total_cost_usd}, each present only when the SDK reports it, plus the SDK's own usage under raw.
error	string | null	When status == "error".

Exit codes

Code	Meaning
0	Success.
1	Runtime error.
2	Missing required arg under --no-interactive / --json.

For run, the exit code is the underlying script's return code on success, 1 if no script was found, or non-zero if --no-interactive would have had to prompt.

Debugging agent failures

Set the RAE_DEBUG environment variable to 1 to print the Python traceback for unexpected agent capture failures to stderr. Normal error messages remain concise by default, and --json / --json-stream stdout remains machine-readable.

Output locations

• ~/.reverse-api/runs/scripts/{run_id}/: permanent storage
• ./scripts/{descriptive_name}/: local copy with a readable name

Caveats

• Generated code runs locally via Claude Code, so review before executing.
• Sites with aggressive bot detection may block capture or require manual interaction.

Development

git clone https://github.com/kalil0321/reverse-api-engineer.git
cd reverse-api-engineer
uv sync
uv run reverse-api-engineer

Build: ./scripts/clean_build.sh. Requires Python 3.11+ and an API key for agent mode. Manual mode additionally needs the [manual] extra plus playwright install chromium.

License

MIT. See LICENSE.