oahelper-mcp

Built by Sirjan Singh

An MCP server that lets Claude Code, Claude Desktop, Cursor, Codex or any other MCP client read your own OAHelper Pro account: company OA question sets, full problem statements, MCQ answer keys, official solutions and interview experiences. Everything it fetches is also saved as plain markdown on your disk, so you can prep offline, in Obsidian, or with an agent that never touches the network again.

It also ships 9 skills for mock OAs, MCQ drills, hint-based coaching, interview briefs, study plans and spaced revision.

The server is one small Python package with a single dependency (the official mcp SDK). HTTP is the standard library.

Quick install

The Claude Code plugin asks for your OAHelper cookie when you enable it and keeps it in secure storage. Its skills show up as /oahelper:oa-mock, /oahelper:mcq-drill and so on.

Read this first. OAHelper's Terms of Service forbid automated access and scraping, and the penalty can be losing your Pro account. This tool is built to be used the way a person uses the site: one company at a time, only when you ask, 2 seconds between requests, a hard daily cap (300 by default), and a disk cache so nothing is fetched twice. It uses your own logged-in browser session, ships no content, and gives you nothing your plan does not already include. Whether to run it is your decision and your risk. Do not use it to crawl the site, and never publish or share what it saves.

What you get

The library layout:

library/<company>/
  index.md            sets, interviews and counts, with links
  sets/<ref>-<name>.md
  questions/<ref>-<title>.md
  interviews/<id>-<role>.md
  images/             only when you ask for them

All of it lives in your data folder: ~/.oahelper by default, the repo folder when you run from a clone, or wherever OAHELPER_HOME points. Inside a clone, library/, cache/ and .env are gitignored. Keep it that way.

Requirements

• uv (recommended), or Python 3.10+ with pip
• An OAHelper account (most content needs Pro)
• Any MCP client that can launch a local stdio server

Add your session

1. Open https://www.oahelper.in in your browser while logged in.

2. Press F12, open the Network tab, reload, and click any request whose URL contains /api/proxy/.

3. Under Request Headers, copy the whole cookie value.

4. Put it in ~/.oahelper/.env (create the folder and file) as one line:

   OAHELPER_COOKIE=<paste here>
   

   .env.example lists the other optional settings. You can instead set OAHELPER_COOKIE as an environment variable in your client's config, and the Claude Code plugin asks for it when you enable it.

The session lasts until you log out on the site. When tools start returning session rejected, repeat these steps. Treat the cookie like a password: do not paste it into a chat with any AI, and do not commit it.

Then ask your agent to run oahelper_status with online=true. You should see session valid | Pro: yes.

Connect it to your agent

With uv installed, every client runs the same command and uv fetches everything else:

uvx --from git+https://github.com/SirjanSingh/oahelper-mcp oahelper-mcp

Claude Code

claude mcp add oahelper -s user -- uvx --from git+https://github.com/SirjanSingh/oahelper-mcp oahelper-mcp

-s user makes it available in every project; use -s local for the current project only. Run claude mcp list to confirm it shows as connected, or /mcp inside a session. Or skip this and install the plugin, which also brings the skills.

Claude Desktop

Edit claude_desktop_config.json (Settings, Developer, Edit Config):

{
  "mcpServers": {
    "oahelper": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/SirjanSingh/oahelper-mcp", "oahelper-mcp"]
    }
  }
}

Restart Claude Desktop.

Cursor

Same JSON as Claude Desktop, in ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project).

VS Code (Copilot agent mode)

.vscode/mcp.json uses a servers key:

{
  "servers": {
    "oahelper": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "git+https://github.com/SirjanSingh/oahelper-mcp", "oahelper-mcp"]
    }
  }
}

OpenAI Codex CLI

codex mcp add oahelper -- uvx --from git+https://github.com/SirjanSingh/oahelper-mcp oahelper-mcp

or in ~/.codex/config.toml:

[mcp_servers.oahelper]
command = "uvx"
args = ["--from", "git+https://github.com/SirjanSingh/oahelper-mcp", "oahelper-mcp"]

Gemini CLI, Windsurf, Cline and others

Any client that reads an mcpServers JSON block takes the Claude Desktop snippet unchanged. Check your client's docs for where its config file lives.

Without uv (from a clone)

git clone https://github.com/SirjanSingh/oahelper-mcp.git
cd oahelper-mcp
python -m venv .venv
.venv/bin/pip install -r requirements.txt     # Windows: .venv\Scripts\pip

Then use /path/to/oahelper-mcp/.venv/bin/python /path/to/oahelper-mcp/server.py as the command in any config above. In this mode .env, library/ and cache/ live in the clone itself.

Using it

Ask in plain language. Some prompts that work well:

• "What OA sets does OAHelper have for Juspay?"
• "Sync Amazon from OAHelper, then tell me the topic mix of their last three OAs."
• "Show me question MjM3OHww with the MCQ answers."
• "Get the official C++ solution for that one."
• "What have people said about Atlassian's interview rounds this year?"
• "What do I already have saved offline?"

A typical prep loop: oahelper_sync a company (call again while it says NOT DONE), then let the agent read library/<company>/index.md and the files it links. Reading the library costs no requests.

Skills

The skills/ folder holds ready-made Agent Skills: plain SKILL.md files that tell the agent how to chain the tools into a workflow. You just ask in plain language and the matching skill kicks in.

oa-mock, mcq-drill and solve-coach log every attempt to library/_progress.md, which study-plan and revise-mistakes read. It stays on your machine with the rest of library/.

Install the skills

Easiest, for Claude Code, Codex, Cursor and other agents that support skills:

npx skills add SirjanSingh/oahelper-mcp

Add --list to see them first, or --skill oa-mock to pick one. In Claude Code, the plugin installs them together with the server.

By hand, for Claude Code in all projects:

cp -r skills/* ~/.claude/skills/

On Windows PowerShell: Copy-Item -Recurse skills\* $HOME\.claude\skills\

For one project only, copy them into that project's .claude/skills/ instead. Restart the session and they show up as /oa-prep, /oa-mock and so on, or trigger on their own from what you ask.

Other agents: the files are plain markdown, so any agent that supports Agent Skills can load them from its skills folder. For one that does not, paste the relevant SKILL.md into its rules or custom instructions.

Configuration

In .env inside the data folder, or as environment variables of the same name (environment wins):

The 2 second gap between requests is MIN_GAP in oahelper_mcp/server.py.

Troubleshooting

How it was built

The site has no public API. The endpoints were found by recording a HAR of normal browsing on a personal account and summarising it with tools/har_endpoints.py, which prints paths, parameter names and response shapes but never header values or response content. If the site changes and something breaks, record a new HAR, run that script on it, and compare.

Releasing (maintainer)

Bump the version in oahelper_mcp/__init__.py, .claude-plugin/plugin.json and both version fields in server.json, commit, then push a tag:

git tag v0.2.2 && git push origin v0.2.2

.github/workflows/release.yml checks the versions match, smoke-tests the wheel, publishes to PyPI (trusted publishing, no token) and then to the MCP Registry (GitHub OIDC, no token).

Author

I am Sirjan Singh. I built this to prep for my own placement OAs.

• GitHub: @SirjanSingh
• LinkedIn: Sirjan Singh

If this saved you prep time, star the repo. It is the only way I find out it helped someone, and it helps other people preparing for OAs find it. Issues and pull requests are welcome.

License

MIT, copyright (c) 2026 Sirjan Singh, for the code in this repository. It grants nothing over OAHelper's content, which belongs to OAHelper and its contributors.

This project is not affiliated with or endorsed by OAHelper.