Stochastly MCP
# Stochastly MCP
<!-- mcp-name: com.stochastly/desktop -->
This small public launcher connects an MCP client to the server shipped with the Stochastly desktop app. The app must be installed, running, and licensed for Atelier or Portfolio. The package contains no Stochastly engine code.
Install with `pip install stochastly-mcp` or run with `uvx stochastly-mcp`. To run this source instead of the PyPI release, install this directory with `pip install .`.
Configure a local stdio server with command `stochastly-mcp` and no arguments:
- Claude Desktop: add `"stochastly": {"command": "stochastly-mcp"}` under `mcpServers` in `claude_desktop_config.json`.
- Cursor: add the same entry under `mcpServers` in `.cursor/mcp.json`.
- VS Code: add `"stochastly": {"type": "stdio", "command": "stochastly-mcp"}` under `servers` in `.vscode/mcp.json`.
The server exposes 29 tools, among them `node_catalog`, `compose_graph`, `backtest_graph`, `red_flag_check`, `adversarial_check`, `portfolio_combine` and `export_paper`. The installed app runs them and checks the licence on every call: `portfolio_combine` needs Portfolio, every other tool Atelier or Portfolio.
Without the app, the launcher lists the same 29 tools (names, descriptions, input schemas and annotations from `stochastly_mcp_data/tools_manifest.json`, generated from the app's server), so a client can see what the server offers. Each description then starts with a note that the desktop app is required, and every call to a listed tool returns a structured answer with `ok: false`, the edition needed and a link to [request access](https://stochastly.com/request-access). A name the server does not list returns the error `Unknown tool`, as the app's server does. The manifest is metadata only.
For a nonstandard installation, set `STOCHASTLY_MCP_SCRIPT` to the installed app's `stochastly_mcp.py`. This is a local path, not a network endpoint.
## Links
- [MCP server overview](https://stochastly.com/product/mcp): what the server does and which clients it supports.
- [MCP server reference](https://stochastly.com/docs/mcp): the 29 tools, their inputs and which ones write.
- [Documentation](https://stochastly.com/docs): the Stochastly desktop app, from data import to research papers.
- [Comparison with other backtesting tools](https://stochastly.com/compare): how Stochastly differs from other backtesting software.
- [Request access](https://stochastly.com/request-access): early access to the desktop app.
TDQS
Scored across 29 tools
Most tools target a distinct resource+action (compose_graph vs backtest_graph vs backtest_status, export_paper vs export_atlas vs export_tearsheet, the v9_* scanner family vs the backtest_* graph family). A few clusters overlap thematically, notably the evaluation tools (red_flag_check, adversarial_check, reality_check_profiles, diagnostics_catalog/explain) and the app-control trio (live_state, set_live_state, pilot_app), but descriptions explicitly cross-reference 'use X instead', resolving most confusion.
Everything is snake_case with a predictable, readable style and useful subsystem prefixes (backtest_, export_, diagnostic_, v9_). The main deviation is that some tools are verb_noun (compose_graph, export_paper, red_flag_check) while others are noun-only (app_info, data_sources, node_catalog, live_state), so the verb pattern isn't fully uniform.
29 tools is on the heavy side and includes clear bloat (set_live_state is explicitly legacy, run_smoke_tests is dev-only), which everyone pays the surface-area cost for. Still, the server spans genuinely different subsystems—graph backtesting, a separate batch scanner, portfolio combination, report export, diagnostics, and app piloting—so the count is defensible rather than gratuitous.
The research lifecycle is well covered: discover nodes/registries/data sources, compose and run graphs, poll status, judge results (red_flag_check, adversarial_check, portfolio_combine), and export papers/tearsheets, plus the parallel v9 scanner surface and app/diagnostics introspection. Minor gaps exist—e.g. no tool to enumerate the day's session run registry directly, and inspection of graph runs depends on a run_id from a prior call—but these are workable.