lse-data-mcp

An unofficial, read-only Model Context Protocol (MCP) server for the London Strategic Edge market-data API.

Status: Alpha. The MCP tool surface may change before the first stable release.

The server lets an MCP client query London Strategic Edge data through the official lse-data Python SDK. It runs locally over standard input/output, uses the API key supplied by the user, returns the upstream JSON-compatible rows, and does not cache or persist responses.

Supported tools

All tools are declared read-only, non-destructive, and idempotent in their MCP metadata.

Every tool returns the same envelope, so a caller can always tell whether it saw the full result:

{
  "rows": [{ "timestamp": "2026-01-02T00:00:00Z", "close": 187.4, "volume": 41230100 }],
  "row_count": 1,
  "truncated": false
}

When the rows would exceed the response budget the server returns the leading rows it can fit, sets "truncated": true, and adds a note explaining how to narrow the request. Rows are never silently dropped.

get_reference groups five discovery endpoints — catalog, datasets, reference, vault_meta, options_underlyings — behind one resource argument, because they take almost no arguments between them. category applies only to catalog and dataset only to datasets; passing either to a resource that ignores it is an error, not a silent no-op, so a grouped tool can never quietly drop a filter you meant. Data tools stay one-to-one with their SDK method, where every argument is always meaningful.

get_reference("catalog") covers 22,000+ instruments, so expect truncated: true unless you filter by category.

Each call defaults to at most 200 rows. The upstream API caps a single interactive call at 5,000 rows; use start and end to request narrower windows.

get_financial_reports defaults to 20 instead, because each row carries a whole statement in its data field and is far larger than a candle or a dividend. Twenty rows is five years of quarterly reports, or twenty years of annual ones.

start and end accept an ISO 8601 date or timestamp (2026-01-01, 2026-01-01T14:30:00Z). Anything else is rejected locally, so a malformed date costs no API call and no quota.

Obtain an API key

1. Visit the official London Strategic Edge data page.
2. Follow the site's prompts to obtain your own API key.
3. Store it with uvx lse-data-mcp login, which prompts without echoing and saves the key to the operating system's own credential store.

Never commit the key to this repository or put a real key in an issue, test, example, or log.

Requirements

• A London Strategic Edge API key
• Either uv, or Python 3.11 or newer

Installation

With uv there is nothing to install: uvx fetches the published package, runs it in a cached environment of its own, and brings its own Python. Store your key, then check it:

uvx lse-data-mcp login
uvx lse-data-mcp status

Whichever command you use here, use the same one in your MCP client below. Mixing uvx with a virtual environment means two different interpreters touch the credential store, which on macOS raises an extra Keychain prompt — see When the server cannot find your key.

Without uv, install the same release from PyPI with pip. Check your interpreter first: macOS ships an older python3 than this project supports, so that command often reports 3.9. Install a supported one with brew install python@3.13 and use it by name; on Windows, use py -3.13.

python3 --version               # must be 3.11 or newer
python3 -m venv .venv           # or python3.13 -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
python -m pip install lse-data-mcp

Activating that environment is what puts lse-data-mcp on your PATH, and a client will need its absolute path rather than the bare uvx command.

To work on the project rather than use it, see CONTRIBUTING.md.

Supplying the API key

Store the key once, in the credential store your operating system already provides:

uvx lse-data-mcp login     # prompts without echoing; nothing is written to a file
uvx lse-data-mcp status    # reports where the key resolves from, without printing it
uvx lse-data-mcp logout    # removes the stored key

Drop the uvx prefix if you installed from source into a virtual environment.

login never accepts the key as a command-line argument, because anything in argv reaches shell history and the process list.

The server resolves its key in this order:

1. the LSE_API_KEY environment variable, when set and non-empty;
2. the credential store written by lse-data-mcp login;
3. otherwise it reports that no key is configured and names both ways to supply one.

The environment wins so that a host injecting the key directly — a container, a CI job, or an MCP client with its own secret manager — stays authoritative over whatever an earlier login left on the machine.

Headless hosts. Secret Service needs a D-Bus session, so a container, an SSH session, or a server install has no credential store to read. Those hosts fall through to LSE_API_KEY rather than failing to start; set it in the environment there.

When the server cannot find your key

lse-data-mcp status distinguishes three outcomes, because they need different fixes:

The third case is what a sandboxed agent runner hits: the server runs in a restricted process, macOS Keychain refuses it, and a key you stored earlier is genuinely there but unreadable. The server reports this as unknown rather than as a missing key, so login is not suggested when re-running it could not help. Grant the host process credential-store access, or pass the key through LSE_API_KEY in the MCP client's environment configuration for that server.

The macOS Keychain prompt

On macOS you may see a dialog like this the first time a given command reads your stored key:

python3.11 wants to use your confidential information stored in "lse-data-mcp" in your keychain. The authenticity of "python3.11" cannot be verified. To allow this, enter the "login" keychain password.

This is expected, and it is macOS asking rather than this server. Keychain records which binary created an entry and asks before letting a different one read it. The dialog names a bare python3.11 because that is the interpreter running the tool — under uvx, a Python that uv manages and that macOS has no signature for.

• Password: your macOS login password, the one you use to unlock the Mac. Not your API key.
• Button: Always Allow records this interpreter against the entry so it stops asking.

The prompt appears at all because the command that stored the key and the command reading it are different programs. Use one or the other consistently and it will not recur:

uvx lse-data-mcp login     # if your MCP client runs `uvx lse-data-mcp`
lse-data-mcp login         # if your client runs a virtual environment's script

It can return after uv upgrades its managed Python, since that is a new binary. If you would rather never see it — on a shared machine, or in an automated environment — set LSE_API_KEY in the MCP client's environment for this server instead, which bypasses the credential store.

.env.example is a reference only. The server deliberately does not load .env files: a .env is plain text on disk, which is what the credential store exists to avoid.

Configuration

An MCP client starts the server for you. To run it by hand — to see a startup error directly, say — use the same command your client does, after storing a key:

uvx lse-data-mcp

From a source checkout, that is lse-data-mcp, or python -m lse_data_mcp to run the package as a module. Nothing is printed on success: the server is waiting to speak JSON-RPC over standard input, so an empty, hanging terminal means it started correctly. Press Ctrl-C to stop it.

MCP client configuration examples

Because the server resolves its own key, no client configuration below contains a secret, and because uvx resolves the package, none of them needs a path.

Claude Code — ~/.claude.json, or run claude mcp add lse-data -- uvx lse-data-mcp:

{
  "mcpServers": {
    "lse-data": {
      "command": "uvx",
      "args": ["lse-data-mcp"]
    }
  }
}

Claude Desktop — claude_desktop_config.json, and Cursor — ~/.cursor/mcp.json for all projects or .cursor/mcp.json for one: same mcpServers object as above.

Codex — ~/.codex/config.toml, which is TOML rather than JSON:

[mcp_servers.lse-data]
command = "uvx"
args = ["lse-data-mcp"]

Restart the client after editing its configuration; MCP servers are spawned at client startup.

Two things to know about command: "uvx". A client launched from the desktop rather than a terminal may not have uvx on its PATH; give the absolute path from which uvx if the server fails to start. And uvx fetches the latest release each time its cache expires, so the server updates itself — pin with ["lse-data-mcp==0.1.0"] if you would rather it did not.

{
  "mcpServers": {
    "lse-data": {
      "command": "/absolute/path/to/.venv/bin/lse-data-mcp"
    }
  }
}

Where a client offers its own secret management and you would rather use it, set LSE_API_KEY through that mechanism; it takes precedence over the stored key. Prefer either of those over a literal key in a configuration file.

Errors and retries

The server converts upstream failures into concise tool errors:

• missing local configuration names both lse-data-mcp login and LSE_API_KEY, without printing any key value;
• HTTP 401 identifies an invalid or expired API key;
• subscription, access, and quota failures explain that the account cannot perform the request;
• HTTP 429 asks the client to wait before retrying;
• timeouts suggest retrying later or increasing LSE_TIMEOUT_SECONDS;
• network and upstream service failures are reported separately.

The server does not automatically retry rate-limited requests. This avoids adding more traffic during an active limit and lets the MCP client decide when to retry.

Known API, data, and subscription limitations

• A tool call returns one interactive page, with a hard maximum of 5,000 rows. This server does not expose bulk history/export jobs.
• A full 5,000-row page is far more JSON than an agent can usefully hold, so the server also caps a result at LSE_MAX_RESPONSE_BYTES and reports the cut through truncated and note. Raise the budget, or page with start and end, when a tool reports truncation.
• Available instruments, fields, history depth, entitlements, quotas, and rate limits are owned by London Strategic Edge and may change. Check the official SDK documentation and your account before relying on a dataset.
• The official SDK states that streaming and downloads share an allowance. Rate-limit or quota exhaustion can therefore be caused by activity outside this MCP process.
• The provider currently documents a free-plan allowance of 10 databank downloads per hour, with up to 1,000,000 rows per download. Those bulk downloads are separate from, and not exposed by, this server.
• The MCP surface is REST-only. Live WebSocket streaming and bulk downloads are out of scope. Every other SDK REST endpoint has a tool, with one deliberate exception. The SDK's economics is a wrapper with no endpoint of its own: without a symbol it returns datasets("economics"), and with one it calls series(symbol, dataset="economics"). Both are already reachable, as get_reference('datasets', dataset='economics') and get_series(symbol, dataset='economics'), so a separate tool would only give an agent two names for one operation.
• get_options returns a live snapshot that refreshes while the market is open, not history, and carries no timestamp of its own. A whole chain on a liquid name runs to thousands of contracts, so filter by expiry, strike or days-to-expiry rather than raising limit.
• get_options_flow covers the trailing week only. Older prints are served as one-minute bars by get_option_candles, whose bars are option premium, not the underlying's price.
• The provider does not document which date field get_financial_reports filters on with start and end — the period end, the fiscal period, or the filing date. Until that is confirmed, prefer period for selecting a fiscal period and treat a date window as approximate.
• get_cot reports a weekly survey, not a live position: the CFTC publishes on Friday for the preceding Tuesday, so the newest row lags the market by several days.
• get_bond_yields returns yields in percent, not prices. They move inversely to price, so a row's high is the day's highest yield and therefore its lowest price.
• A stock split rebases historical prices and share counts. Check get_splits before comparing any per-share figure across a window that contains one.
• Market data may be delayed, incomplete, corrected, or unavailable. It is not investment advice.

Repository and data-safety policy

This repository must contain integration code and synthetic test data only. Do not commit:

• API keys, tokens, credentials, or populated .env files;
• downloaded proprietary datasets;
• real API responses containing restricted data;
• a public or hosted proxy that serves data using the maintainer's credentials.

Every user runs the adapter with their own key. Tests use recording fakes and synthetic rows; they must never call the live API or consume an account allowance. Local MCP clients may retain tool results in conversation history or logs, so users must configure those clients consistently with the provider's data terms.

Development checks

ruff format --check .
ruff check .
mypy src tests
pytest

GitHub Actions runs all four checks on Python 3.11, 3.12, and 3.13 for pushes and pull requests.

Project structure

src/lse_data_mcp/
├── __init__.py     # package metadata
├── __main__.py     # python -m entry point
├── cli.py          # command line: run the server, or manage the stored key
├── client.py       # lazy upstream SDK client
├── config.py       # key resolution and environment configuration
├── credentials.py  # operating-system credential store
├── server.py       # FastMCP server and tool registration
└── tools.py        # read-only market-data tools

Data rights and unofficial-project disclaimer

The MIT license covers this integration code only. It does not grant rights to London Strategic Edge data, APIs, SDKs, names, or trademarks. Review the provider's terms before using or retaining returned data; in particular, do not redistribute or resell data unless the provider expressly permits it.

This is an independent community project. It is not affiliated with, endorsed by, sponsored by, or maintained by London Strategic Edge.

License

MIT