rigor-mcp
rigor
为 AI 代理提供经过验证的统计推断。
LLM 在背诵统计知识方面还算不错,但在可靠地执行上却很糟糕——
t 统计量或所需样本量这类数字往往是从训练数据中回忆出来的,而不是计算并核验出来的。rigor 就是替代方案:
经典假设检验、效应量、功效/样本量计算以及多重比较校正,全部从零计算,
并作为附有引用且经过假设核查的答案返回。
它构建为一个 MCP 服务器:对当前 MCP 生态系统的扫描(Context7 用于编程文档, 多个物理/工程/化学/地理服务器,甚至 Bentley 的 STAAD 集成)发现, 统计学/实验设计是为数不多的、尚无人覆盖的常见代理需求之一。
统计部分本身(rigor/distributions.py、inference.py、
effect_size.py、power.py、corrections.py)是纯标准库实现,没有依赖。
整个包确实依赖于官方 mcp SDK,因为 MCP 服务器是它所提供功能的一等组成部分,
而不是附加组件——请参阅安装。
安装
pip install rigor-mcp(PyPI 发行版名为 rigor-mcp,因为单纯的 rigor 已被一个无关的包占用;
可导入的包和 CLI 命令仍然都叫 rigor。)这样你会同时获得两个控制台命令:
rigor(CLI)和 rigor-mcp(MCP 服务器)——有意设计为一次安装即可,
无需配置额外依赖,因为 uvx rigor-mcp(大多数 MCP 客户端实际调用它的方式)
无法请求任何附加依赖。
Related MCP server: evalgate
包含内容
rigor/distributions.py—— t、卡方和 F 分布,在标准库上从零构建(正则化的不完全伽马/贝塔函数), 通过精确的闭式恒等式进行验证(t(1) = Cauchy、chi2(2) = 缩放指数分布、t² = F(1, df)), 而不是依赖照搬现成的转录。rigor/inference.py—— 单样本/双样本和配对 t 检验、单比例/双比例 z 检验、 卡方拟合优度与独立性检验、单因素方差分析。每个都返回一个TestResult: 统计量、自由度、双尾 p 值、置信区间、引用以及假设警告 (例如小样本时对正态性的依赖、期望单元格计数过低)。rigor/effect_size.py—— Cohen's d、Hedges' g、Cohen's h、Cramér's V。rigor/power.py—— 双样本 t 检验和双比例 z 检验的功效与所需样本量。 两个方向(给定 n 求功效;给定功效求 n)在构造上就是彼此的精确数值逆运算 (在同一底层功效函数上使用二分法),并通过 Cohen (1988) d=0.5/α=.05/power=.80 教科书参考案例(n≈64)进行了合理性检查。rigor/corrections.py—— Bonferroni 和 Benjamini-Hochberg(FDR)多重比较校正。rigor/cli.py—— 覆盖上述所有功能的 CLI(仓库根目录下的rigor.py是一个轻量垫片, 因此python3 rigor.py ...也能在普通检出目录中运行,无需安装任何东西)。rigor/mcp_server.py—— 一个 MCP 工具包装器,向任何 MCP 客户端(Claude Code、Claude Desktop 等) 暴露全部 17 个操作。通过 stdio 对真实客户端进行了端到端冒烟测试——工具发现以及代表性调用 均与已知参考值对照核验,包括完整往返仍然落在 Cohen (1988) 案例的 n=63。
用法
CLI(安装后):
rigor ttest one-sample --data 5.1,4.9,5.3,5.0,4.8,5.2 --mu0 5.0
rigor power ttest-2samp --effect-size 0.5 --power 0.8
rigor --help # full list of subcommands (ttest, ztest, chi2, anova, effect-size, power, correct)或者直接从检出目录运行,无需安装任何东西:
python3 rigor.py ttest one-sample --data 5.1,4.9,5.3,5.0,4.8,5.2 --mu0 5.0MCP 服务器,通过 stdio(本地客户端如 Claude Code 所期望的传输方式):
pip install rigor-mcp
rigor-mcp或从检出目录运行:pip install mcp && python3 -m rigor.mcp_server。
将其注册到 Claude Code:
claude mcp add rigor -- rigor-mcp(或者,从检出目录运行:claude mcp add rigor -- python3 -m rigor.mcp_server,
在此仓库的根目录下运行,或使用绝对模块路径)。要用 MCP Inspector 进行交互式探索,
请将其作为脚本运行,而不是作为已安装的命令——这意味着需要手动将包根目录添加到路径中,
因为 Inspector 会直接导入该文件:
pip install "mcp[cli]"
PYTHONPATH=. mcp dev rigor/mcp_server.py一个传输层边界情况,已处理
cohens_d 对于零方差样本会正确地返回 +inf/-inf(依据其自身记录的约定),
但非有限浮点数在 MCP 的结构化内容中会被序列化为 JSON null——这曾导致工具自身的
number 类型输出模式校验失败,并使调用崩溃。MCP 的 cohens_d 工具现在返回
{"value": float | null, "warnings": [...]},而不是一个裸浮点数,因此这种情况会被明确报告
(值为 null,并有一条指明方向的警告),而不是直接崩溃。这里所有其他数值工具的输出都是有界的,
对于有效输入始终为有限值,因此这种处理方式仅适用于 cohens_d。
测试
python3 -m unittest discover -s tests -v70 个测试:65 个直接测试统计功能;5 个像真实 MCP 客户端那样启动 mcp_server.py,
并通过传输层检查结果(如果未安装 mcp,会自动跳过)。
许可证
MIT——参见 LICENSE。
Available Tools
37 toolsbenjamini_hochberg_correctionARead-onlyIdempotent
Adjust a batch of p-values for multiple comparisons, controlling the false discovery rate. Less conservative than Bonferroni; the standard choice when testing many hypotheses at once.
Inputs: p_values (required), alpha.
bonferroni_correctionARead-onlyIdempotent
Adjust a batch of p-values for multiple comparisons, controlling the family-wise error rate. Conservative; use when any false positive among the batch is costly.
Inputs: p_values (required), alpha.
chi_square_goodness_of_fitARead-onlyIdempotent
Test whether observed category counts match an expected distribution -- e.g. "are these six days-of-week signup counts evenly distributed, or skewed towards weekends?" Returns the chi-squared statistic, degrees of freedom (len-1), p-value, a citation, and a warning if any expected count is below 5 (the usual threshold below which this approximation gets unreliable).
Inputs: expected (required), observed (required), alpha.
chi_square_independenceARead-onlyIdempotent
Test whether the row and column variables of a contingency table are independent (e.g. "does group membership relate to outcome?"). Returns the chi-squared statistic, degrees of freedom, p-value, a citation, and a warning if any expected cell count is below 5 (consider cramers_v afterwards for effect size).
Inputs: table (required), alpha.
cohens_dARead-onlyIdempotent
Standardized mean difference between two samples (pooled SD). Use alongside two_sample_t_test, which tells you whether a difference is significant but not how large it is. Rough guidance: ~0.2 small, ~0.5 medium, ~0.8 large -- context-dependent. Returns {"value": float or null, "warnings": [...]}. value is null only when both samples have zero variance and unequal means, where the effect size is mathematically infinite -- see the warning for which direction, and report the raw mean difference instead in that case.
Inputs: a (required), b (required).
cohens_hARead-onlyIdempotent
Effect size for a difference between two proportions (Cohen, 1988), via the arcsine-square-root transform -- more appropriate than a raw percentage-point difference since it stabilizes variance across the full [0, 1] range. p1 and p2 are interchangeable (the sign of the result just indicates direction); use alongside two_proportion_z_test, which tells you whether a difference is significant but not how large it is. Returns a float (can be negative); rough guidance: ~0.2 small, ~0.5 medium, ~0.8 large.
Inputs: p1 (required), p2 (required). Returns: result
cramers_vARead-onlyIdempotent
Effect size for a chi-squared test of independence (Cramer, 1946), normalized to [0, 1] regardless of table shape so it's comparable across tables of different sizes, unlike the raw chi-squared statistic. Call after chi_square_independence, passing its statistic and the same table's n/rows/cols. Returns a float in [0, 1]; rough guidance for a 2x2 table: ~0.1 small, ~0.3 medium, ~0.5 large -- the threshold shifts for larger tables.
Inputs: n (required), cols (required), rows (required), chi2_statistic (required). Returns: result
eta_squaredARead-onlyIdempotent
Effect size for a one-way ANOVA: proportion of total variance explained by group membership. Use alongside one_way_anova, which tells you whether groups differ but not how much of the variance that accounts for. Rough guidance: ~0.01 small, ~0.06 medium, ~0.14 large. Biased upward for small samples -- prefer omega_squared when that matters. Returns a float in [0, 1].
Inputs: groups (required). Returns: result
fisher_exact_testARead-onlyIdempotent
Test whether the row and column variables of a 2x2 contingency
table are independent -- exact (via the hypergeometric distribution
over all tables with the same margins), unlike
chi_square_independence's chi-squared approximation. Use this
instead whenever chi_square_independence warns an expected cell
count is below 5, or whenever the sample is small. 2x2 tables only.
Returns the sample odds ratio as statistic (can be inf/0 for a
zero cell), a two-tailed p-value, a citation, and warnings.
Inputs: table (required), alpha.
kruskal_wallisARead-onlyIdempotent
The non-parametric alternative to one_way_anova -- use when that test's own small-df warning makes a normal-theory result suspect. Tests whether all groups are drawn from the same distribution, by ranking the combined data rather than assuming normal populations. A significant result means at least one group differs, not which one -- same caveat as one_way_anova.
Inputs: groups (required), alpha.
levene_testARead-onlyIdempotent
Test whether two or more groups have equal population variances (homogeneity of variance) -- use this to decide equal_var for two_sample_t_test, or to sanity-check one_way_anova's equal-variance assumption. Uses the Brown-Forsythe variant (deviations from each group's median), more robust to non-normal data than the original mean-based Levene's test. Returns the same shape as one_way_anova (it's computed as one internally, on absolute deviations from each group's median).
Inputs: groups (required), alpha.
mann_whitney_uARead-onlyIdempotent
The non-parametric alternative to two_sample_t_test -- use when that test's own small-n warning makes a normal-theory result suspect, or the data is ordinal/skewed. Tests whether values from sample a are systematically larger or smaller than values from sample b, by ranking the combined data rather than assuming normal populations. statistic is U for sample a; pair with rank_biserial_correlation for a standardized effect size. Returns the same result shape as the parametric tests (statistic, p_value, citation, warnings).
Inputs: a (required), b (required), alpha.
mcnemar_exact_testARead-onlyIdempotent
Exact version of mcnemar_test: an exact binomial test (p=0.5) on
the discordant pairs instead of the chi-squared approximation -- the
small-sample-safe alternative mcnemar_test's own warning points to,
the same relationship fisher_exact_test has to
chi_square_independence. statistic is b-c (the raw
discordant-pair imbalance); the p-value is two-tailed.
Inputs: table (required), alpha.
mcnemar_testARead-onlyIdempotent
Test whether two paired proportions are equal -- e.g. the same subjects' yes/no answers before and after an intervention, or two raters' calls on the same items. Use this instead of two_proportion_z_test whenever the "two groups" are actually the same subjects measured twice; two_proportion_z_test assumes independent groups and gets the standard error wrong for paired data. Yates continuity-corrected chi-squared, 1 df -- use mcnemar_exact_test instead when there are few discordant pairs (this warns when there are). Returns that chi-squared statistic, a p-value, and a confidence interval for the difference in marginal proportions.
Inputs: table (required), alpha.
naive_peeking_inflationARead-onlyIdempotent
Demonstrates, by simulation, why sequential_two_sample_mean_test / sequential_two_proportion_test exist: the actual false-positive rate of checking an ordinary fixed-sample test (two_sample_t_test, two_proportion_z_test, ...) after every new observation and stopping the first time it clears alpha, versus the alpha actually intended. Call this to show a skeptical stakeholder concretely what "just peek at the dashboard and stop early" costs before recommending the always-valid alternative. Returns the estimated true false-positive rate, its Monte Carlo standard error, and a citation.
Inputs: n_looks (required), alpha, trials.
omega_squaredARead-onlyIdempotent
Effect size for a one-way ANOVA, less biased than eta_squared for small samples since it subtracts out the variance explained by chance alone. Use alongside one_way_anova. Can be slightly negative when the true effect is near zero -- that's expected, not an error.
Inputs: groups (required). Returns: result
one_proportion_z_testARead-onlyIdempotent
Test whether an observed proportion (successes out of n) differs from a hypothesized proportion p0 -- e.g. "is this coin fair (p0=0.5) given 55 heads in 100 flips?" Uses the normal approximation, which degrades for small n or p0 near 0 or 1; a warning is included when that assumption looks shaky. Returns the z-statistic, two-tailed p-value, a confidence interval for the true proportion, a citation, and warnings.
Inputs: n (required), p0 (required), successes (required), alpha.
one_sample_t_testARead-onlyIdempotent
Test whether a sample's mean differs from a hypothesized value mu0. Returns the t-statistic, degrees of freedom, two-tailed p-value, a confidence interval for the mean, and any assumption warnings.
Inputs: mu0 (required), data (required), alpha.
one_way_anovaARead-onlyIdempotent
Test whether three or more independent groups have different means -- e.g. comparing average order value across three marketing channels. A significant result means at least one group differs from the others, not which one -- follow up with pairwise two_sample_t_test calls (correcting for multiple comparisons via bonferroni_correction or benjamini_hochberg_correction) to find which. Returns the F-statistic, between/within degrees of freedom, p-value, a citation, and a warning if within-group df is small.
Inputs: groups (required), alpha.
paired_t_testARead-onlyIdempotent
Test whether the mean difference between paired observations (e.g. before/after measurements on the same subjects, or matched pairs) is zero. a[i] and b[i] must be the two measurements of the same pair -- use two_sample_t_test instead if the two samples are independent (different subjects in each group). Returns the t-statistic, degrees of freedom (n-1), two-tailed p-value, a confidence interval for the mean difference, a citation, and assumption warnings.
Inputs: a (required), b (required), alpha.
pairwise_group_comparisonsARead-onlyIdempotent
Run every pairwise comparison across 2+ groups and correct for multiple comparisons in one call, instead of orchestrating k*(k-1)/2 separate two_sample_t_test/mann_whitney_u calls plus a separate correction call by hand -- and forgetting the correction is one of the most common real mistakes this package exists to prevent. The natural follow-up after a significant one_way_anova/kruskal_wallis result: pass the same groups here to find which group(s) differ, not just whether any do. Returns every pair's statistic, raw p-value, whether it's still significant after correction, and an effect size, plus the correction method's citation and warnings.
Inputs: groups (required), test, alpha, labels, +2 more.
pearson_correlationARead-onlyIdempotent
Test for a linear association between two paired variables -- e.g. "does hours studied predict test score?" statistic is r itself (in [-1, 1]), not a t-statistic. Returns r, df (n-2), a two-tailed p-value (H0: r=0), a confidence interval for r via the Fisher z-transform, a citation, and warnings. Use spearman_correlation instead if the relationship may be monotonic but not linear, or if outliers shouldn't dominate the result. Use simple_linear_regression instead for the actual slope (units of y per unit of x), not just the strength of association.
Inputs: x (required), y (required), alpha.
power_for_one_sample_t_testARead-onlyIdempotent
Statistical power to detect a given Cohen's d with n observations, using a one-sample (or paired) t-test. Use for paired_t_test too -- it's a one-sample t-test on the differences, so the same power formula applies. Use sample_size_for_one_sample_t_test instead to solve for n given a target power. Returns a float in [alpha, 1].
Inputs: n (required), effect_size_d (required), alpha. Returns: result
power_for_two_proportion_testARead-onlyIdempotent
Statistical power to detect a difference between two proportions (e.g. two conversion rates) with n_per_group observations in each group, using a two-proportion z-test. p1 and p2 are interchangeable (only their difference matters) -- e.g. current vs. new conversion rate. Use sample_size_for_two_proportion_test instead to solve for n given a target power. Returns a float in [alpha, 1].
Inputs: p1 (required), p2 (required), n_per_group (required), alpha. Returns: result
power_for_two_sample_t_testARead-onlyIdempotent
Statistical power to detect a given Cohen's d with n_per_group observations per group, using a two-sample t-test. Power is the probability of correctly detecting a real effect of this size at the given alpha; a design with low power means a non-significant result would be inconclusive rather than good evidence the effect doesn't exist. Use sample_size_for_two_sample_t_test instead to solve for n given a target power. Returns a float in [alpha, 1].
Inputs: n_per_group (required), effect_size_d (required), alpha. Returns: result
rank_biserial_correlationARead-onlyIdempotent
Effect size for a Mann-Whitney U test. Call after mann_whitney_u, passing its statistic and the two sample sizes. Positive means sample 1's values tend to exceed sample 2's; negative means the reverse; 0 is no tendency either way. Returns a float in [-1, 1]; rough guidance mirrors Cohen's d: ~0.1 small, ~0.3 medium, ~0.5 large.
Inputs: n1 (required), n2 (required), u1_statistic (required). Returns: result
recommend_testARead-onlyIdempotent
Not sure which rigor tool fits your question? Answer a few characteristics of the data and get back which tool to call, why, what to call instead if this test's assumptions look shaky, and what to run alongside it (an effect size, a power calculation, a natural follow-up). Every test in this package already documents this guidance in its own docstring for the sibling comparisons it knows about -- this tool exists so you don't have to have already read every other tool's docstring to find the one relevant cross-reference. Pure decision logic, no statistics computed here.
Inputs: outcome_type (required), paired, n_groups, small_or_skewed, +3 more.
sample_size_for_one_sample_t_testARead-onlyIdempotent
How many observations are needed to detect a given Cohen's d with a one-sample (or paired) t-test at the target power. Use for paired_t_test too -- a paired t-test is a one-sample t-test on the differences, so the same power formula applies. Returns a continuous value and a rounded-up integer to actually use.
Inputs: effect_size_d (required), alpha, target_power.
sample_size_for_two_proportion_testARead-onlyIdempotent
How many observations per group are needed to detect a difference between two proportions (e.g. conversion rates) at the target power. p1 and p2 are interchangeable -- only their difference matters.
Inputs: p1 (required), p2 (required), alpha, target_power.
sample_size_for_two_sample_t_testARead-onlyIdempotent
How many observations per group are needed to detect a given Cohen's d with a two-sample t-test at the target power. Returns a continuous value and a rounded-up integer to actually use.
Inputs: effect_size_d (required), alpha, target_power.
sequential_two_proportion_testARead-onlyIdempotent
Always-valid test of whether two proportions (e.g. two conversion rates in a live A/B test) differ, safe to call again after every new observation in either group -- the sequential-monitoring counterpart to two_proportion_z_test. Use this instead whenever the result will be checked more than once before the experiment ends, which is the normal case for a live dashboard rather than a one-shot analysis. p1 and p2 are interchangeable (only their difference matters). tau does not need to be exact -- reuse the minimum-detectable-effect you'd otherwise plug into sample_size_for_two_proportion_test.
Inputs: n1 (required), n2 (required), tau (required), successes1 (required), +2 more.
sequential_two_sample_mean_testARead-onlyIdempotent
Always-valid test of whether two groups' means differ, safe to call again after every new observation in either group -- unlike two_sample_t_test, which needs a sample size decided in advance and gives no such guarantee if checked repeatedly and stopped at the first significant look (that repeated-checking failure mode is exactly what inflates false positives; see naive_peeking_inflation for a demonstration). Use this instead of two_sample_t_test whenever a result will be (or already has been) checked more than once as data accumulates, e.g. monitoring a live experiment. Returns the current effect estimate, its standard error, the mixture likelihood ratio and always-valid p-value, and assumption warnings. tau does not need to be exact -- reuse the minimum-detectable-effect you'd otherwise plug into sample_size_for_two_sample_t_test.
Inputs: a (required), b (required), tau (required), alpha, +1 more.
simple_linear_regressionARead-onlyIdempotent
Fit y = intercept + slope * x by ordinary least squares -- single predictor only. Reports the slope (change in y per unit of x), the intercept, R^2 (proportion of y's variance explained by x), and a significance test + confidence interval for the slope (H0: slope=0). Use pearson_correlation instead if you only need the strength of a linear association, not its actual units/magnitude.
Inputs: x (required), y (required), alpha.
spearman_correlationARead-onlyIdempotent
Test for a monotonic association between two paired variables, via the Pearson correlation of their ranks -- doesn't assume linearity and is far less sensitive to outliers' exact magnitude than pearson_correlation. Same return shape as pearson_correlation (statistic is rho itself, in [-1, 1]).
Inputs: x (required), y (required), alpha.
two_proportion_z_testARead-onlyIdempotent
Test whether two independent proportions differ -- the standard test behind comparing conversion rates between two groups (e.g. an A/B test). Returns the z-statistic, two-tailed p-value, a confidence interval for the difference in proportions, a citation, and warnings.
Inputs: n1 (required), n2 (required), successes1 (required), successes2 (required), +1 more.
two_sample_t_testARead-onlyIdempotent
Test whether two independent samples have different means. Defaults to Welch's t-test (does not assume equal variances); pass equal_var=true for the classic pooled-variance test.
Inputs: a (required), b (required), alpha, equal_var.
wilcoxon_signed_rankARead-onlyIdempotent
The non-parametric alternative to paired_t_test -- use when that test's own small-n warning makes a normal-theory result suspect. Tests whether the median of the paired differences is zero, by ranking the absolute differences rather than assuming they're normally distributed. Pairs with a zero difference are dropped (and counted in a warning), the standard procedure. statistic is T = min(W+, W-).
Inputs: a (required), b (required), alpha.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 37 tools
Every tool targets a distinct statistical procedure or design (e.g., paired vs. independent proportions, fixed-sample vs. sequential means, chi-square vs. exact test), and the descriptions explicitly cross-reference sibling tools to prevent misselection. Even the paired power/sample-size tools are clearly differentiated as inverse operations.
All tool names use consistent snake_case and follow a predictable pattern: tests are named after the procedure (e.g., two_sample_t_test, mann_whitney_u), effect sizes after the statistic, and power/sample-size tools share the power_for_/sample_size_for_ prefix. The few non-test utilities (recommend_test, naive_peeking_inflation) are still stylistically consistent.
37 tools is well beyond the 25+ threshold for a heavy surface, even though each tool is individually useful. The inclusion of recommend_test and pairwise_group_comparisons mitigates the burden, but the set would benefit from consolidation or a more focused scope.
The suite covers the common lifecycle of a statistical analysis: assumption checks, parametric/nonparametric tests, effect sizes, power/sample-size planning, multiple-comparison corrections, and pairwise follow-ups. Minor gaps remain, such as no power/sample-size calculations for ANOVA or chi-square and no explicit normality test, but agents can work around these.
Maintenance
Related MCP Connectors
Exact statistics & probability: distributions, hypothesis tests, CIs, Bayesian updates, regression.
Valid and reliable data engineering and statistical analysis without hallucinations.
Verify before you cite: URL checks, quotes, page-to-Markdown, metadata, schema, links, archives.
111A statistical analyst in your AI chat: bring data and a question, get a citable, re-runnable report.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables comprehensive statistical analysis including descriptive statistics, hypothesis testing, regression, and more via a FastMCP-based API.3-
- AlicenseAqualityCmaintenanceEval-integrity statistics for AI benchmark claims — multiple-testing correction, power/MDE for model gaps, judge-bias and leaderboard-rank checks. Catches a benchmark number that won't survive a second look.9MIT
- AlicenseAqualityBmaintenanceEnables AI assistants to run verified statistical and analytics computations—including hypothesis tests, A/B testing, descriptive summaries, and sample-size planning—directly on local CSV, TSV, Excel, and Parquet files of any size. Results include effect sizes, confidence intervals, plain-English interpretations, and assumption warnings, with optional independent checks.2373 npmMIT
- AlicenseAqualityBmaintenanceEnables MCP clients and AI agents to compute sample sizes and statistical power for the G*Power t-test family through six tools — listing and describing tests, running power analyses, computing effect sizes, generating power curves, and reporting engine provenance — with every result returned as the engine's verbatim JSON.6MIT