Evalsmith Mcp

evalsmith answers four questions that come up every time a retriever, prompt or agent changes:

1. Is retrieval good enough? recall@k, precision@k, nDCG@k, hit rate and MRR from qrels and a ranked run.
2. Which configuration is better, and is the gap real? 2-3 configurations ranked with seeded paired-bootstrap confidence intervals.
3. Did the agent call the right tools? Trace vs. expected calls, strict or lax order, argument matchers.
4. Did the prompt change break anything? Golden-set checks on baseline vs. candidate outputs.

Everything is deterministic and runs offline. An LLM judge exists for a second opinion, but it is off by default and never decides the verdict. Results come out as a terminal table, Markdown, JSON, JUnit XML or a self-contained HTML report, and the GitHub Action turns them into a merge gate.

[evalsmith run: retrieval metrics, config comparison, tool-call check, prompt regression and a failing CI gate]

Interfaces over the same core

flowchart LR    subgraph Clients        A[MCP client<br/>desktop app, IDE, agent]        B[Remote MCP client<br/>or service]        C[Terminal / CI job]        D[GitHub workflow]    end    subgraph evalsmith        S1[MCP server<br/>stdio]        S2[MCP server<br/>Streamable HTTP, stateless]        S3[CLI<br/>evalsmith run]        S4[GitHub Action<br/>uses: evalsmith-mcp@v0]        CORE[(src/core<br/>metrics, bootstrap, matchers,<br/>checks, gates, reports)]    end    A --> S1 --> CORE    B --> S2 --> CORE    C --> S3 --> CORE    D --> S4 --> CORE    S3 -. exit code + JUnit + HTML .-> C    S4 -. outputs + job summary .-> D

All four call the same functions in src/core. The MCP tools are thin adapters: parse input with Zod, call the core, return structuredContent validated against an output schema.

MCP surface

flowchart TB    subgraph Tools        T1[evaluate_retrieval]        T2[compare_configs]        T3[check_tool_calls]        T4[detect_prompt_regression]        T5[generate_report]        T6[run_suite]    end    subgraph Resources        R1["evalsmith://examples/{name}<br/>12 synthetic datasets + suite"]    end    subgraph Prompts        P1[compare_rag_configs]        P2[audit_agent_trace]        P3[review_prompt_change]    end    T1 -- structuredContent --> T5    T2 -- structuredContent --> T5    T3 -- structuredContent --> T5    T4 -- structuredContent --> T5    classDef tool fill:#ffedd5,stroke:#c2410c,color:#431407    classDef res fill:#e0f2fe,stroke:#0369a1,color:#082f49    classDef prom fill:#ede9fe,stroke:#6d28d9,color:#2e1065    class T1,T2,T3,T4,T5,T6 tool    class R1 res    class P1,P2,P3 prom

Tools

Every tool declares an input and an output Zod schema, returns structuredContent plus a JSON text block, and is annotated readOnly and non-destructive. Four tools are also idempotent and closed-world. detect_prompt_regression and run_suite are declared open-world and non-idempotent: annotations are per tool, and with the judge enabled they call an external model whose answers can vary. With the judge off (the default) they behave like the others.

run_suite runs the same suites as evalsmith run, so a client can evaluate data that lives in files instead of pasting it into the conversation. It only reads inside the workspace root (EVALSMITH_ROOT, the directory the server started in by default), following symlinks before it checks, and runs the bundled example suite when no config is given.

Bad input (unknown metric, duplicate config names, qrels with no relevant documents) comes back as an InvalidParams tool error with a message the model can act on. A judge endpoint that is down or returns an HTTP error is an InternalError.

Resources

evalsmith://examples/{name} lists and serves the synthetic datasets in examples/datasets, so a client can try every tool before wiring in its own data:

Prompts

A conversation with an MCP client

sequenceDiagram    autonumber    actor User    participant Host as MCP client (LLM host)    participant ES as evalsmith (stdio)    participant Core as src/core
    User->>Host: "Is the reranker worth it on our help-center set?"    Host->>ES: initialize, tools/list, resources/list    ES-->>Host: 6 tools + JSON Schemas, 13 resources    Host->>ES: resources/read evalsmith://examples/rag-qrels (+ 3 runs)    ES-->>Host: qrels and runs as JSON    Host->>ES: tools/call compare_configs {metric: "ndcg@5", configs: [...]}    ES->>Core: compareConfigs(qrels, configs, "ndcg@5", {seed: 42})    Core-->>ES: ranking, intervals, paired delta    ES-->>Host: structuredContent {winner: "hybrid+rerank", significant: true, ...}    Host->>ES: tools/call generate_report {format: "markdown", comparison}    ES-->>Host: Markdown report    Host-->>User: "Yes: +0.153 nDCG@5 over hybrid, 95% CI [0.069, 0.251]."

Metrics

For a query $q$ with relevant set $R_q$ (documents with grade $> 0$) and ranking $D_q$, with $D_q^{k}$ its first $k$ documents:

\mathrm{Recall@}k = \frac{|R_q \cap D_q^{k}|}{|R_q|}\qquad\mathrm{Precision@}k = \frac{|R_q \cap D_q^{k}|}{k}\qquad\mathrm{HitRate@}k = \mathbb{1}\left[\,R_q \cap D_q^{k} \neq \varnothing\,\right]

\mathrm{RR}_q = \frac{1}{\text{rank of the first relevant document}}\qquad\mathrm{MRR} = \frac{1}{|Q|}\sum_{q \in Q} \mathrm{RR}_q

\mathrm{DCG@}k = \sum_{i=1}^{k} \frac{\mathrm{rel}_i}{\log_2(i+1)}\qquad\mathrm{nDCG@}k = \frac{\mathrm{DCG@}k}{\mathrm{IDCG@}k}

$\mathrm{rel}_i$ is the graded judgement of the document at rank $i$ (linear gain, as in trec_eval), and IDCG is the DCG of the judgements sorted in descending order. Reported values are means over judged queries.

Conventions, all covered by tests:

• Queries whose judgements are all 0 are skipped and listed in skippedQueries.
• Judged queries missing from the run score 0 and are listed in missingFromRun.
• Precision divides by $k$, not by the ranking length.
• Repeated document ids in a ranking keep only their first position, as in trec_eval, so a run cannot score above 1 by repeating a relevant document.
• Cut-offs start at 1: recall@0 is rejected.

Worked example (tests/core/retrieval.test.ts): judgements a=2, b=1, c=0, ranking d, a, e, b, c. DCG@5 $= 2/\log_2 3 + 1/\log_2 5 = 1.69254$, IDCG@5 $= 2/\log_2 2 + 1/\log_2 3 = 2.63093$, so nDCG@5 $= 0.64332$; recall@3 $= 0.5$, precision@5 $= 0.4$, RR $= 0.5$.

Bootstrap. compare_configs resamples queries with replacement, using the same indices for every configuration (paired), samples times, from a seeded PRNG (mulberry32). Intervals are percentiles of the resampled means. The lead is significant when the interval of winner minus runner-up excludes zero. The same seed gives the same intervals on every machine.

Tool-call matchers

{  "tool": "create_refund",  "args": {    "invoice_id": { "type": "exact", "value": "INV-2041" },    "reason": { "type": "any" },    "metadata": { "type": "subset", "value": { "ticket": "T-5532" } },    "to": { "type": "regex", "pattern": "@acme-robotics\\.example$" }  }}

strict requires the trace to be exactly the expected calls in order; calls past the expected sequence fail the check and appear as failed test cases in JUnit. lax ignores order and pairs expected calls with trace calls by maximum bipartite matching (augmenting paths), so a loose matcher never steals the only call a stricter one accepts. Extra calls are reported without failing. Every matcher, exact included, fails when the argument is missing.

Prompt checks

exact (trimmed by default), regex, contains (case-insensitive by default), json_schema (the output is parsed and validated with Ajv) and length (min / max characters). A regression is a case that passes on the baseline and fails on the candidate. Missing outputs count as failures.

The optional judge sends both outputs to any OpenAI-compatible /chat/completions endpoint, Ollama included, at temperature 0. To cancel position bias each case is judged twice, with baseline as answer A and then as answer B. A side wins only when both orderings agree; a disagreement counts as a tie. That costs two requests per case. Its votes are reported next to the verdict.

Quickstart

npx -y evalsmith-mcp                       # MCP server on stdionpm install -g evalsmith-mcp               # or install the `evalsmith` CLI and `evalsmith-mcp` serverdocker build -t evalsmith-mcp . && docker run --rm -i evalsmith-mcp   # or run the server in a container

To run the bundled examples and the tests from source:

git clone https://github.com/diegoperezg7/evalsmith-mcp.gitcd evalsmith-mcpnpm ci            # also builds dist/ through the prepare scriptnpx playwright install chromium   # the HTML report tests open it in a real browsernpm test          # metrics, traces, MCP in-memory + HTTP, CLI exit codes, HTML report, action
node dist/cli/index.js run examples/evalsmith.yaml

Connect an MCP client (stdio)

[Install in Cursor] [Install in VS Code]

From a terminal:

claude mcp add evalsmith -- npx -y evalsmith-mcp     # Claude Codecodex mcp add evalsmith -- npx -y evalsmith-mcp      # Codex CLIcode --add-mcp '{"name":"evalsmith","command":"npx","args":["-y","evalsmith-mcp"]}'   # VS Code

Other desktop MCP clients read an mcpServers block like this (examples/mcp-client-config.json):

{  "mcpServers": {    "evalsmith": {      "command": "npx",      "args": ["-y", "evalsmith-mcp"],      "env": { "LOG_LEVEL": "warn" }    }  }}

Restart the client, then try the compare_rag_configs prompt or ask it to read evalsmith://examples/agent-trace and check it.

Streamable HTTP

EVALSMITH_TRANSPORT=http EVALSMITH_HTTP_PORT=3333 node dist/mcp/server.jscurl -s localhost:3333/healthz                  # {"status":"ok","version":"0.1.0"}npx @modelcontextprotocol/inspector             # connect to http://127.0.0.1:3333/mcp

CLI and CI

evalsmith run <config> [--fail-under metric=value ...] [--format table|json|markdown|html]                       [--report file.{md,json,xml,html}] [--report-format markdown|json|junit|html]evalsmith doctor [config] [--judge]

Gates accept recall@k, precision@k, ndcg@k, hit_rate@k, mrr, tool_calls (score of check_tool_calls) and prompt_pass_rate. Gates in the YAML file apply too; a command-line gate with the same name overrides it. A prompt regression or a failed tool-call check fails the run even without gates.

Real output

node dist/cli/index.js run examples/evalsmith.yaml on the bundled datasets (examples/evalsmith.yaml):

evalsmith · ledgerly-support-assistant
Retrieval (12 queries, MRR 0.8403)┌────┬────────┬───────────┬────────┬──────────┐│  k │ recall │ precision │   nDCG │ hit rate │├────┼────────┼───────────┼────────┼──────────┤│  1 │ 0.4583 │    0.7500 │ 0.6667 │   0.7500 ││  3 │ 0.7500 │    0.4167 │ 0.7442 │   0.9167 ││  5 │ 0.9167 │    0.3000 │ 0.8059 │   1.0000 ││ 10 │ 1.0000 │    0.1667 │ 0.8285 │   1.0000 │└────┴────────┴───────────┴────────┴──────────┘
Comparison on ndcg@5 (paired bootstrap, 2000 samples, seed 42)┌───┬───────────────┬────────┬──────────────────┐│ # │ config        │ ndcg@5 │ 95% CI           │├───┼───────────────┼────────┼──────────────────┤│ 1 │ hybrid+rerank │ 0.9588 │ [0.8950, 1.0000] ││ 2 │ hybrid        │ 0.8059 │ [0.6875, 0.9141] ││ 3 │ bm25          │ 0.5596 │ [0.3782, 0.7354] │└───┴───────────────┴────────┴──────────────────┘hybrid+rerank wins on ndcg@5 (0.9588), +0.1528 over hybrid; 95% interval of the gap [0.0692, 0.2506] excludes zero.
Agent tool calls (strict)┌──────────────────┬─────────┬────────┬───────┐│ expected         │ matched │ status │ notes │├──────────────────┼─────────┼────────┼───────┤│ search_customers │ #1      │ pass   │       ││ get_invoice      │ #2      │ pass   │       ││ create_refund    │ #3      │ pass   │       ││ send_email       │ #4      │ pass   │       │└──────────────────┴─────────┴────────┴───────┘All 4 expected call(s) matched (strict order).
Prompt regression┌───────────────────────┬──────────┬───────────┬────────────────────┐│ case                  │ baseline │ candidate │ candidate failures │├───────────────────────┼──────────┼───────────┼────────────────────┤│ classify-intent       │ pass     │ pass      │                    ││ extract-invoice-id    │ pass     │ pass      │                    ││ refund-policy-summary │ fail     │ pass      │                    ││ due-date-format       │ pass     │ pass      │                    ││ refuse-card-number    │ pass     │ pass      │                    ││ refund-payload        │ pass     │ pass      │                    │└───────────────────────┴──────────┴───────────┴────────────────────┘No regressions across 6 case(s). Pass rate 83.3% -> 100.0%.
Gates┌────────────┬────────┬────────┬────────┐│ metric     │    min │ actual │ status │├────────────┼────────┼────────┼────────┤│ recall@5   │ 0.8000 │ 0.9167 │ pass   ││ ndcg@5     │ 0.7500 │ 0.8059 │ pass   ││ tool_calls │ 1.0000 │ 1.0000 │ pass   │└────────────┴────────┴────────┴────────┘
PASS

Raising the bar makes the same run fail the build:

$ node dist/cli/index.js run examples/evalsmith.yaml --fail-under recall@5=0.95 --fail-under mrr=0.9 --report report.xml...Gates┌────────────┬────────┬────────┬────────┐│ metric     │    min │ actual │ status │├────────────┼────────┼────────┼────────┤│ recall@5   │ 0.9500 │ 0.9167 │ FAIL   ││ ndcg@5     │ 0.7500 │ 0.8059 │ pass   ││ tool_calls │ 1.0000 │ 1.0000 │ pass   ││ mrr        │ 0.9000 │ 0.8403 │ FAIL   │└────────────┴────────┴────────┴────────┘
FAIL · 2 gate(s) below thresholdReport written to report.xml$ echo $?1

HTML report

--report report.html (or --format html on stdout) writes the whole run to one HTML file that opens offline and follows the system light or dark mode:

• the verdict, and a card per gate with its value, threshold and status;
• retrieval metrics per k as bars with a bootstrap interval over the queries, and gate thresholds marked on them;
• the configuration comparison with each interval, the winner and the gap to the runner-up;
• expected vs. actual tool calls, with the matched pairs linked and mismatches flagged;
• baseline vs. candidate per golden case, with the failing checks one click away.

[evalsmith HTML report in light mode: FAIL verdict, gate cards and retrieval metrics per k with intervals]

Dataset text (case ids, tool names, failure messages) is embedded as JSON in a non-executable <script type="application/json"> block and rendered with textContent only. A Content-Security-Policy allows no network access and only the page's own script, by hash. Tests open a report full of markup in Chromium and check that nothing from it runs. npm run docs:screenshots regenerates the pictures above.

GitHub Action

- uses: actions/checkout@v5- uses: diegoperezg7/evalsmith-mcp@v0  id: evals  with:    config: evals/evalsmith.yaml    fail-under: |      recall@5=0.8      ndcg@10=0.7      tool_calls=1    html-report: evalsmith-report.html

The step fails when a gate is below its threshold, the prompt candidate regresses or the tool calls do not match, and writes the Markdown report with the verdict to the job summary. It runs the same core as the CLI from a bundled node24 script, with no install step and no shell. examples/github-workflow.yml is a complete workflow that also keeps the HTML report as an artifact and publishes the JUnit XML as a check run.

Outputs of a section the suite does not have are empty. Dataset text printed to the log is wrapped in ::stop-commands::, so a case id or tool name cannot issue workflow commands. The judge reads EVALSMITH_JUDGE_* from the step environment, as the CLI does.

Without the action, the CLI does the same from any CI:

npx evalsmith run evals/evalsmith.yaml --fail-under recall@5=0.8 --format markdown --report evalsmith-report.xml >> "$GITHUB_STEP_SUMMARY"

Suite file

name: ledgerly-support-assistantretrieval:  qrels: datasets/rag/qrels.json # { queryId: { docId: grade } }  run: datasets/rag/runs/hybrid.json # { queryId: [docId, ...] }  k: [1, 3, 5, 10]compare:  metric: ndcg@5  configs:    - { name: bm25, run: datasets/rag/runs/bm25.json }    - { name: hybrid, run: datasets/rag/runs/hybrid.json }    - { name: hybrid+rerank, run: datasets/rag/runs/hybrid-rerank.json }  bootstrap: { samples: 2000, seed: 42, confidence: 0.95 }agent:  trace: datasets/agent/trace.json  expected: datasets/agent/expected.json  order: strictprompts:  golden: datasets/prompts/golden.json  baseline: datasets/prompts/baseline.json  candidate: datasets/prompts/candidate.json  judge: { enabled: false }gates:  recall@5: 0.8  ndcg@5: 0.75  tool_calls: 1.0

Every section is optional. Paths are relative to the YAML file.

Pairs with ragcompass

ragcompass profiles a corpus and recommends how to configure the RAG pipeline: parser, chunking, top-k, hybrid search, reranking. evalsmith checks whether that configuration holds up:

flowchart LR    A[Documents] --> B[ragcompass analyze<br/>recommended configs]    B --> C[Index 2-3 candidate configs]    C --> D[Run the eval queries<br/>save rankings]    D --> E[evalsmith compare_configs<br/>winner + intervals]    E --> F[evalsmith run in CI<br/>--fail-under gates]

Design decisions

• Deterministic first. Metrics, matchers and checks are pure functions with hand-computed tests. A model is never needed to get a verdict, so results do not drift between runs and no paid API key is required.
• Judge off by default. When enabled it goes through any OpenAI-compatible endpoint (local Ollama works), and its votes sit next to the verdict instead of changing it.
• Intervals, not just means. Twelve queries can make a 0.03 lead look decisive. A paired bootstrap with a fixed seed makes the uncertainty visible and reproducible.
• Stateless HTTP done properly. A stateless StreamableHTTPServerTransport cannot be reused across requests, so every POST gets its own server and transport. Registration is cheap, and a test makes two consecutive calls over real HTTP to prove it.
• stdout belongs to the protocol. pino writes to stderr and ESLint forbids console in src/, so no stray byte breaks the stdio transport. The CLI keeps stdout for results only.
• One schema, three uses. The Zod schemas in src/core/schemas.ts validate MCP input, generate the JSON Schemas clients see, and type the core.

Project layout

src/  core/         metrics, bootstrap, tool-call matchers, prompt checks, judge, gates, reports (incl. HTML), suite runner  mcp/          server (stdio + HTTP), tools/, resources/, prompts/  cli/          evalsmith run | doctor, table/json/markdown/html output  action/       GitHub Action entry point, bundled into dist/action  shared/       config (Zod), pino logger on stderr, MCP errorstests/  core/         hand-computed metrics, bootstrap, matchers, checks, reports, HTML in Chromium  mcp/          InMemoryTransport client and real Streamable HTTP client  cli/          exit codes 0 / 1 / 2, JUnit file, doctorexamples/  datasets/     synthetic RAG corpus + qrels + 3 runs, agent trace, golden prompts  evalsmith.yaml, github-workflow.yml, mcp-client-config.json

npm run check     # typecheck + eslint + prettier + vitestnpm run inspect   # MCP Inspector against the built stdio server

Releases come from version tags: release.yml publishes to npm with provenance, to the MCP Registry and to GitHub Releases, and the action is published to the Marketplace from the release. docs/RELEASING.md has the steps, including the action bundle and the v0 tag. Changes are listed in CHANGELOG.md.

License

MIT