rti-mcp

An MCP server for querying your own applications on India's RTI Online portal — status, filed text, replies and reply PDFs — without re-entering an OTP and two captchas for every lookup.

Ask your assistant "which of my RTIs are overdue?" instead of clicking through View History one application at a time.

> Which of my RTI requests are more than 30 days overdue?

  14 of your 146 pending requests are past the Act's 30-day deadline.
  The oldest is DGDOR/R/E/21/00133, filed 1,844 days ago…

> What happened to the NIMHANS one?

  NIMNS/R/E/26/00220 — REQUEST DISPOSED OF as on 30/07/2026. A reply PDF
  is available; want me to download it?

Contents

• Why this works

• Install

• Get your session URL

• Register with your MCP client

• Tools

• Configuration

• How the portal behaves

• Troubleshooting

• Security

• Disclaimer

Related MCP server: MANIT ERP MCP Server

Why this works

Logging in to RTI Online and clicking View History lands you on a citizen_view_history.php URL whose emailchk, cellchk and urletoken parameters are server-side encrypted blobs. That URL authenticates itself — it keeps working from a plain HTTP client with no cookies carried over from the browser, and it stays valid for a long time.

This server takes that one URL and walks the rest of the portal from it. You pay the OTP + captcha cost once, whenever the URL eventually stops working.

It is read-only. It only ever reads your own account, and only what the portal already shows you when you are logged in. It cannot file, appeal, edit or pay. Anyone holding your URL can read the same data, so treat it like a password — see Security.

Install

Requires Python 3.10+.

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install rti-mcp

Install into a virtual environment rather than system Python: the next section asks for an absolute interpreter path, and a venv is what makes that path stable.

git clone https://github.com/gouthamganeshm/rti-mcp.git
cd rti-mcp
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

Or the unreleased main, without a working copy:

pip install "git+https://github.com/gouthamganeshm/rti-mcp.git"

Point your MCP config at an absolute interpreter path, never a bare python. A bare python resolves to whatever environment happens to be active when the client spawns the server, and the server vanishes from the list with a ModuleNotFoundError the moment that differs from the environment holding the dependencies. Get the right path with:

python -c "import sys; print(sys.executable)"

Get your session URL

1. Open https://rtionline.gov.in and log in with your OTP and captcha.

2. Click View History.

3. Copy the entire URL from the address bar. It looks like:

   https://rtionline.gov.in/request/citizen_view_history.php?emailchk=…&cellchk=…&urletoken=…

Then hand it to the server, either by asking your assistant —

"Set my RTI session URL to https://rtionline.gov.in/request/citizen_view_history.php?emailchk=…"

— which calls rti_set_session_url and stores it in ~/.rti-mcp/config.json, or by exporting it before the client starts:

export RTI_HISTORY_URL="https://rtionline.gov.in/request/citizen_view_history.php?emailchk=…"

The stored config file wins over the environment variable, so rti_set_session_url can refresh an expired URL at runtime without touching your MCP client config or restarting anything.

Register with your MCP client

Each client keeps its own registry — registering with one does not populate another's list.

claude mcp add rti-online -s user -- /absolute/path/to/.venv/bin/python -m rti_mcp

On Windows the interpreter sits elsewhere in the venv, so pass that path instead:

claude mcp add rti-online -s user -- C:\path\to\.venv\Scripts\python.exe -m rti_mcp

Re-pointing an existing entry means removing it first — claude mcp add will not overwrite one that is already registered:

claude mcp remove rti-online -s user

Or a project .mcp.json (key: mcpServers):

{
  "mcpServers": {
    "rti-online": {
      "type": "stdio",
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["-m", "rti_mcp"]
    }
  }
}

Verify with claude mcp list, or /mcp inside a session.

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "rti-online": {
      "command": "C:\\path\\to\\.venv\\Scripts\\python.exe",
      "args": ["-m", "rti_mcp"]
    }
  }
}

Restart Claude Desktop afterwards.

.vscode/mcp.json — note the key is servers, not mcpServers:

{
  "servers": {
    "rti-online": {
      "type": "stdio",
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["-m", "rti_mcp"]
    }
  }
}

The server will not appear in the Extensions sidebar; that is expected.

The server speaks stdio and is also installed as a console script:

rti-mcp

Equivalent to python -m rti_mcp. Point any MCP-capable client at either.

Tools

Things to ask for once it is wired up:

• "What's the status of NIMNS/R/E/26/00220?"

• "Which of my RTIs are more than 30 days overdue?"

• "Anything change on my applications today?"

• "Download the reply for the NIMHANS one."

• "Export everything to a spreadsheet."

Configuration

Every list-returning tool also takes refresh=true to bypass the cache for one call.

How the portal behaves

Worth knowing, because the constraints shaped the code:

• No pagination. A bucket's list page carries every row — several hundred registered requests in one ~300 KB response. One fetch gets everything.

• registered is the superset. Its count equals disposed + pending, so rti_search and rti_export_csv fetch one list per category, not three.

• Detail links are single-use-ish. Every list fetch mints fresh regId/token params, valid only while that fetch is the most recent successful navigation. Fetching a different list invalidates the previous page's links.

• A 403 poisons the session. Once one request 403s, the next one fails too, whatever it is. Re-walking seed → list clears it.

client.py encodes these rules: detail links are never cached to disk, several details can be read off one live list, and a 403 triggers an automatic re-walk from the seed URL. Requests are paced ~0.6 s apart to stay polite to a government server.

One known quirk, surfaced rather than hidden: the dashboard's counts sometimes run a row or two ahead of its own list pages (e.g. it says 146 pending while the pending page lists 144). That is the portal's inconsistency, not a parsing gap — the row parser matches the served table exactly, and any row it fails to recognise is reported by rti_session_status instead of being dropped.

Troubleshooting

The server does not appear in the client's tool list. Almost always the interpreter path. Run /absolute/path/to/python -m rti_mcp in a terminal: if it fails with ModuleNotFoundError: No module named 'rti_mcp', the package is installed into a different environment than the one your config names.

"The RTI Online session URL is no longer valid." It expired. Log in again, open View History, copy the URL, and pass it to rti_set_session_url. Nothing else needs changing.

Results look stale. Lists are cached for 15 minutes. Pass refresh=true, or call rti_clear_cache.

A tool reports unparsed_rows. The portal changed its registration-number format and some applications are missing from results. Please open an issue with the shape of the number that failed — not your real one.

Everything 403s. The session got poisoned mid-walk. The client recovers automatically; if it persists, rti_clear_cache then retry.

Security

~/.rti-mcp/config.json holds a URL that grants read access to your entire RTI account. Don't commit it, don't paste it into an issue, and don't put it in a screenshot. There is no logout — revoking it means waiting for the portal to expire it.

Full details, and how to report a vulnerability, in SECURITY.md.

Disclaimer

Not affiliated with, endorsed by, or connected to the Government of India, the Department of Personnel and Training, or the RTI Online portal.

This is an unofficial client that parses HTML the portal was not designed to serve to programs, so it can break whenever the portal changes. It reads only the account whose session URL you supply. Use it for your own applications; don't point it at the portal at large, and don't remove the request pacing.

The 30-day figure rti_overdue uses is the ordinary deadline under §7(1) of the RTI Act, 2005 — shorter and longer periods apply in some cases (48 hours where life or liberty is concerned, 35 or 40 days when routed through an APIO or a third party is involved). Check the Act before relying on a date.

Use this at your own risk.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. The test suite runs entirely offline against synthetic fixtures, so you can work on the parsers without an RTI account.

License

MIT © Goutham Ganesh M H