Skip to main content
Glama

choson-mcp

An MCP server for the Chosŏn conjoint API: upload survey data, fit, read results and run the simulators from Claude (Claude Code, Claude Desktop) or any MCP client. It runs on your machine and calls https://quantecarlo.com/api/v1 with your API key. Paid Chosŏn plans.

Set up

  1. Create an API key: sign in at https://quantecarlo.com/app/, click API, Create key.

  2. Add the server.

Claude Code:

claude mcp add choson -e CHOSON_API_KEY=chs_... -- uvx choson-mcp

Claude Desktop (claude_desktop_config.json):

{"mcpServers": {"choson": {"command": "uvx", "args": ["choson-mcp"], "env": {"CHOSON_API_KEY": "chs_..."}}}}

Until the package is on PyPI, install from GitHub instead: replace uvx choson-mcp with uvx --from git+https://github.com/sign-of-fourier/choson-mcp choson-mcp (in the Desktop config, args ["--from", "git+https://github.com/sign-of-fourier/choson-mcp", "choson-mcp"]).

The API itself is documented at https://quantecarlo.com/api.md.

Related MCP server: Local Workspace Orchestrator

Tools

whoami, list_projects, create_project, get_project, upload_data (a local file), get_dataset, set_design, quote_fit (hardware, estimate and the 2-hour maximum, nothing queued), start_fit, fit_status, wait_for_fit, cancel_fit, get_result (whole or one section), download (per-respondent CSV, workbook, Excel simulator), simulate_shares, simulate_removal (diversion), price_sweep, elasticities, forecast, api_reference (the full API reference) and call_api (any other /v1 route: range, recommend, test-next, optimize, segments, the menu, MaxDiff, volumetric and claims tools, survey design).

The server asks the assistant to show you the hardware, estimate and maximum from quote_fit and get your go-ahead before starting a fit. Every fit is stopped at 2 hours by the platform.

Environment

variable

default

CHOSON_API_KEY

required

CHOSON_API_URL

https://quantecarlo.com/api/v1

Available Tools

21 tools
api_referenceA

The full Chosŏn API reference: data formats for every module, fit settings, the results JSON, every simulator and search with its fields, errors. Read it before call_api or when a format is unclear.

Inputs: no named parameters. Returns: result

call_apiB

Any other /v1 route (see api_reference): searches (range, recommend, test-next; queued, poll with GET on the same path), optimize, segments, MBC / MaxDiff / volumetric / concept tools, survey design. path is relative to /v1, e.g. "/fits/abc123/range". JSON only; use upload_data and download for files.

Inputs: path (required), method (required), body.

cancel_fitC

Cancel a queued fit at once or a running one within seconds.

Inputs: fit_id (required).

create_projectB

Create a project. kind: cbc, mbc, maxdiff, pricing, concept or volumetric.

Inputs: name (required), kind.

downloadA

Save a fit's file to the local disk. what: respondents.csv (per-respondent estimates), workbook.xlsx (the study workbook) or excel-simulator.xlsx (CBC, needs the Chosŏn Excel add-in).

Inputs: what (required), fit_id (required), save_to (required). Returns: result

elasticitiesC

Own- and cross-price elasticities for a shelf, with 95% ranges.

Inputs: fit_id (required), products (required).

fit_statusB

A fit's status (queued, running, done, failed, cancelled), progress stage, error and timestamps.

Inputs: fit_id (required).

forecastC

Units, revenue and profit with 95% ranges, the change against a base shelf and the chance of clearing a hurdle. request: products, occasions, prices, unit_costs, own (indexes), optional base and hurdle (see api_reference).

Inputs: fit_id (required), request (required).

get_datasetD

A dataset: validation report, summary, design.

Inputs: dataset_id (required).

get_projectC

A project with its datasets and fits.

Inputs: project_id (required).

get_resultA

The results JSON of a finished fit. Large; pass section to get one part: derived (part-worths, importances, WTP), params (mean tastes, taste spread, error shares), fit (log-likelihood, holdout hit rate, warnings), baselines (comparison models), meta (settings, design). Every estimate is {est, se, lo, hi} with a 95% range.

Inputs: fit_id (required), section.

list_projectsA

All projects on the account (id, name, kind).

Inputs: no named parameters.

price_sweepC

Shares as one product's (0-based index) price moves across the tested range (or the given prices).

Inputs: fit_id (required), product (required), products (required), prices.

quote_fitA

What a fit with these settings would run on, without queuing it: compute (cpu, modal_cpu, t4, h100), estimated_seconds, max_seconds (every fit is stopped at 7,200 s = 2 hours) and any notes. Same settings as start_fit. Show the user this before starting a fit.

Inputs: dataset_id (required), settings.

set_designB

CBC: replace the dataset's design (attributes with name, levels, kind categorical / linear; none; price) and re-run the checks. Start from the design get_dataset returns and edit it.

Inputs: design (required), dataset_id (required).

simulate_removalB

Remove one product (0-based index into products): the shares after, and where its buyers go (diversion to each remaining product and to none), with 95% ranges. The substitution question.

Inputs: fit_id (required), remove (required), products (required), group.

simulate_sharesB

Share of preference for a shelf, with 95% ranges. products: up to 60, each {attribute: level} with levels exactly as in the design (price any number in the tested range). group: a covariate group, if the fit has them.

Inputs: fit_id (required), products (required), group.

start_fitA

Queue a fit. settings (all optional, see api_reference): taste (none / diagonal / full), error_shares (list of attributes), holdout (last / none), hb_baseline, constraints, covariates, classes, compute (auto / cpu / t4 / h100). Returns the fit with compute (the hardware chosen) and progress.estimated_seconds. Every fit stops at 2 hours. Call quote_fit first, show the user the hardware, the estimate and the 2-hour maximum, and get their go-ahead.

Inputs: dataset_id (required), settings.

upload_dataA

Upload a responses file from the local disk (CSV in the module's long format, or a Sawtooth .CHO for CBC; see api_reference for formats). Optional design JSON for CBC. Returns the dataset: its id, the validation report (errors must be fixed before fitting; read warnings) and, for CBC, the design (check design_inferred).

Inputs: project_id (required), responses_path (required), design_path.

wait_for_fitA

Poll a fit until it is done, failed or cancelled, or max_seconds pass (at most 600; call again to keep waiting). Returns the last status.

Inputs: fit_id (required), max_seconds.

whoamiA

The account's plan and its limits (respondents, products, tasks, queued jobs) and GPU minutes this month.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3/5.0

    Scored across 21 tools

    Disambiguation4/5

    Tools target clearly distinct operations: project CRUD, dataset handling, fit lifecycle (quote/start/status/wait/result/cancel), and four separate simulation queries. Minor overlap exists between forecast and simulate_shares (both take a product shelf) and call_api acts as a catch-all that can reach any route, but the descriptions delineate these well.

    Naming Consistency4/5

    Most names follow a predictable verb_noun pattern (create_project, upload_data, start_fit, simulate_shares). A few deviate as bare nouns or noun_verb (fit_status, elasticities, forecast, price_sweep, whoami), with fit_status notably breaking the get_* pattern used by get_result/get_project.

    Tool Count4/5

    21 tools is on the heavy side but justified by a genuinely complex domain spanning multiple modules (CBC, MaxDiff, pricing, concept, volumetric) with a full fit lifecycle plus simulation suite. Each tool maps to a real operation rather than redundant surface.

    Completeness4/5

    Covers the full workflow from project creation through data upload, fitting, results, and simulation, plus docs and an escape hatch. Gaps remain (no delete/update project, no list_fits, no dataset deletion), but call_api lets agents reach missing routes as a workaround.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Connects Claude to the WhatConverts API, enabling management of leads, accounts, profiles, tracking numbers, and recordings through natural language.
      1
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables testing and interacting with any REST API endpoints directly from MCP clients such as Claude and Cline, with support for multiple HTTP methods, authentication, custom headers, and response-size management.
      5 npm
      ISC
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables Claude to query client and document data in plain English with cited answers, enforce per-client access, safely file incoming documents, and record audited actions through MCP tools.
      MIT