io.github.hidai25/evalview-mcp

The open-source behavior regression gate for AI agents.
Think Playwright, but for tool-calling and multi-turn AI agents.

Your agent can still return 200 and be wrong. A model or provider update can change tool choice, skip a clarification, or degrade output quality without changing your code or breaking a health check. EvalView catches those silent regressions before users do — and gives you the loop to investigate them, grade the confidence, and broadcast the verdict to your team.

You don't need frontier-lab resources to run a serious agent regression loop. EvalView gives solo devs, startups, and small AI teams the same core discipline: snapshot behavior, detect drift, classify changes, and review or heal them safely.

Traditional tests tell you if your agent is up. EvalView tells you if it still behaves correctly. It tracks drift across outputs, tools, model IDs, and runtime fingerprints with graded confidence — not a binary alarm — so you can tell "the provider changed" from "my system regressed."

30-second live demo.

Most eval tools stop at detect and compare. EvalView helps you classify changes, inspect drift, and auto-heal the safe cases.

• Catch silent regressions that normal tests miss
• Separate provider/model drift from real system regressions
• Auto-heal flaky failures with retries, review gates, and audit logs

Built for frontier-lab rigor, startup-team practicality:

• targeted behavior runs instead of giant always-on eval suites
• deterministic diffs first, LLM judgment where it adds signal
• faster loops from change -> eval -> review -> ship

How we run EvalView with this operating model →

  ✓ login-flow           PASSED
  ⚠ refund-request       TOOLS_CHANGED
      - lookup_order → check_policy → process_refund
      + lookup_order → check_policy → process_refund → escalate_to_human
  ✗ billing-dispute      REGRESSION  -30 pts
      Score: 85 → 55  Output similarity: 35%

The money screen is the one-line verdict that lands under every check — a single ship/don't-ship decision derived from the diff, quarantine state, cost delta, and drift confidence:

─────────────────────────────────────────────
 VERDICT: 🛑 BLOCK RELEASE
─────────────────────────────────────────────

  • 1 regression: billing-dispute
  • 1 test changed behavior: refund-request
  • Cost up 14% vs baseline

Likely cause & next actions:

  1. Rerun statistically to distinguish flake from real drift
     (high severity, high confidence)
     → evalview check --statistical 5

  2. Review tool descriptions for: escalate_to_human
     (high severity, high confidence)
     Tool selection changed — usually a prompt edit nudged the model
     → evalview replay refund-request --trace
     → evalview golden update refund-request   # if the new path is correct

Four tiers: SAFE_TO_SHIP, SHIP_WITH_QUARANTINE, INVESTIGATE, BLOCK_RELEASE. The verdict is part of --json output, the PR comment, and the cloud ship page — CLI, CI, and dashboard all tell the same story.

Quick Start

pip install evalview

evalview init        # Detect agent, auto-configure profile + starter suite
evalview snapshot    # Save current behavior as baseline
evalview check       # Catch regressions after every change

That's it. Three commands to regression-test any AI agent. init auto-detects your agent type (chat, tool-use, multi-step, RAG, coding) and configures the right evaluators, thresholds, and assertions.

After check, the investigative loop:

evalview since                             # 2-second brief: what's changed since your last run
evalview progress --since yesterday        # delta report: what improved/regressed
evalview drift                             # per-test sparklines + incident markers
evalview slack-digest --webhook $SLACK     # post the daily verdict to your team

These four commands turn a red ✗ into an answer — is it real drift, a known flake, or a provider update? — before anyone opens Slack. since is the habit anchor (daily open-the-terminal glance); progress answers "did my changes help?" with a worth-a-commit gate; drift visualizes the trend; slack-digest broadcasts the verdict. See Daily Workflow →.

Catch silent drift in closed models

Worried that claude-opus-4-5 might behave differently next week without warning? evalview model-check runs a zero-judge canary suite directly against the provider and tells you exactly when the model has drifted — no agent required, no calibration.

evalview model-check --model claude-opus-4-5-20251101   # first run saves baseline
evalview model-check --model claude-opus-4-5-20251101   # next week, detects any change

See the full model-check section →

curl -fsSL https://raw.githubusercontent.com/hidai25/eval-view/main/install.sh | bash

evalview demo        # See regression detection live (~30 seconds, no API key)

git clone https://github.com/hidai25/evalview-support-automation-template
cd evalview-support-automation-template
make run

evalview generate --agent http://localhost:8000           # Generate tests from a live agent
evalview capture --agent http://localhost:8000/invoke      # Capture real user flows (runs assertion wizard after)
evalview capture --agent http://localhost:8000/invoke --multi-turn  # Multi-turn conversation as one test
evalview generate --from-log traffic.jsonl                # Generate from existing logs
evalview init --profile rag                               # Override auto-detected agent profile

Daily Workflow

Detection is only the first step. EvalView gives you the full investigative loop — so when a test goes red, you can answer "is it real drift, a known flake, or a provider update?" in four commands, before anyone opens Slack.

Open the terminal — evalview since

╭──────────────────────────── Since your last check ─────────────────────────────╮
│                                                                                │
│   95%  pass rate across 14 runs                                                │
│   ⚠  2 soft change(s)                                                          │
│   ✨ improved: search_cases, summarize_thread                                  │
│                                                                                │
╰────────────────────────────────────────────────────────────────────────────────╯

Drift sparklines (most-declining first)
  ▇▆▅▄▃▂▁▂  billing-dispute

🔥 Streak: 6 days of clean merges

One thing to look at first:
  → evalview replay billing-dispute

evalview since is the 2-second habit brief — one hero number, one concern, one action. It reuses Week 1's fingerprinted history, so the "since your last check" window is accurate whether you ran it 10 minutes ago or 4 days ago. Night owls at 2am and daily shippers at 9am see the same command; the label adapts. It's the command that goes in your .zshrc so evalview since fires when you open the terminal, before the espresso machine is warm.

Morning — evalview progress --since yesterday

✨ 3 test(s) now passing that weren't
⚠  1 test(s) regressed

Improved:
  + refund-flow
  + order-lookup (at a4f2e91)

Regressed:
  − billing-dispute

Output similarity: 85.20% → 87.50% ↑ +2.30%

Worth a commit:
  ✓ refund-flow (high confidence)
    → evalview golden update refund-flow

A "worth a commit" gate (3+ consecutive passes) keeps you from celebrating flakes at 2am.

Triage — evalview drift billing-dispute

Test             │ Trend        │ Samples │ Slope  │ First → Last │ Status
─────────────────┼──────────────┼─────────┼────────┼──────────────┼────────
billing-dispute  │            ! │ 20      │ -1.5%  │ 90% → 80%    │ declining
                 │ ▇▆▅▄▃▂▁▂▁▂▁ │         │        │              │

Most concerning: billing-dispute — slope -1.50% per check over 20 samples
  → evalview replay billing-dispute --trace

Unicode sparklines + OLS slope + incident markers (!) show when the test flipped. Drift is graded insufficient_history / stable / low / medium / high — not a binary alarm.

Verdict — evalview check --statistical 5

When the verdict layer returns INVESTIGATE, a stability-replay recommendation is auto-injected at position #0 of the action list, surviving the severity sort so you never miss it.

Quarantine — evalview quarantine list

Test              │ Owner   │ Age  │ Flaky │ Trend │ Status    │ Reason
──────────────────┼─────────┼──────┼───────┼───────┼───────────┼──────────────
race-condition    │ @hidai  │ 12d  │ 4     │ ↘     │ ⏸ active │ race condition
db-timeout        │ @jane   │ 45d  │ 8     │ ↗     │ ⏰ STALE  │ db timeout

⏰ 1 entry stale — review overdue.
   Either fix the underlying flake or remove from quarantine:
   evalview quarantine remove db-timeout

Known-flaky tests don't block CI — but staleness tracking, owner tags, and a flaky-count trend glyph keep the list honest. Governance built in, not a dumping ground.

Broadcast — evalview slack-digest --webhook $SLACK_WEBHOOK

📊 EvalView digest — yesterday
🟢 95% pass rate across 47 runs

Drift
  ▇▆▅ billing-dispute

⏰ Stale quarantine
  1 overdue
  • db-timeout — @jane — 45d

🎯 Next: evalview check --fail-on REGRESSION

✓ Digest posted to Slack.

Stdlib-only Block Kit post (zero new deps). Fails soft on bad webhooks. Ends with one actionable command your team can copy-paste from the channel.

The loop closes: detection → investigation → graded verdict → quarantine governance → broadcast. You wake up, run progress, triage with drift, confirm with check --statistical, and the team sees the digest before standup. That's the morning ritual — reach for it before the espresso machine warms up.

Why EvalView?

Use LangSmith for observability. Use Braintrust for scoring. Use EvalView for regression gating.

Detailed comparisons →

What It Catches

Model / Runtime Change Detection

EvalView does more than compare model_id.

• Declared model change: adapter-reported model changed from baseline
• Runtime fingerprint change: observed model labels in the trace changed, even when the top-level model name is missing
• Coordinated drift: multiple tests shift together in the same check run, which often points to a silent provider rollout or runtime change — now graded low / medium / high via DriftTracker.classify_drift, not a binary alarm

When detected, evalview check surfaces a run-level signal with a classification (declared or suspected), confidence level, and evidence from fingerprints, retries, and affected tests.

If the new behavior is correct, rerun evalview snapshot to accept the updated baseline.

Four scoring layers — the first two are free and offline:

Score Breakdown
  Tools 100% ×30%    Output 42/100 ×50%    Sequence ✓ ×20%    = 54/100
  ↑ tools were fine   ↑ this is the problem

CI/CD Integration

Block broken agents in every PR. One step — PR comments, artifacts, and job summary are automatic.

# .github/workflows/evalview.yml — copy this, add your secret, done
name: EvalView Agent Check
on: [pull_request, push]

jobs:
  agent-check:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write
    steps:
      - uses: actions/checkout@v4

      - name: Check for agent regressions
        uses: hidai25/eval-view@main
        with:
          openai-api-key: ${{ secrets.OPENAI_API_KEY }}

## ✅ EvalView: PASSED

| Metric | Value |
|--------|-------|
| Tests | 5/5 unchanged (100%) |

---
*Generated by EvalView*

## ❌ EvalView: REGRESSION

> **Alerts**
> - 💸 Cost spike: $0.02 → $0.08 (+300%)
> - 🤖 Model changed: gpt-5.4 → gpt-5.4-mini

| Metric | Value |
|--------|-------|
| Tests | 3/5 unchanged (60%) |
| Regressions | 1 |
| Tools Changed | 1 |

### Changes from Baseline
- ❌ **search-flow**: score -15.0, 1 tool change(s)
- ⚠️ **create-flow**: 1 tool change(s)

Common options: strict: 'true' | fail-on: 'REGRESSION,TOOLS_CHANGED' | mode: 'run' | filter: 'my-test'

Also works with pre-push hooks (evalview install-hooks) and status badges (evalview badge).

Full CI/CD guide →

Watch Mode

Leave it running while you code. Every file save triggers a regression check.

evalview watch                          # Watch current dir, check on change
evalview watch --quick                  # No LLM judge — $0, sub-second
evalview watch --test "refund-flow"     # Only check one test

╭─────────────────────────── EvalView Watch ────────────────────────────╮
│   Watching   .                                                        │
│   Tests      all in tests/                                            │
│   Mode       quick (no judge, $0)                                     │
╰───────────────────────────────────────────────────────────────────────╯

14:32:07  Change detected: src/agent.py

╭──────────────────────────── Scorecard ────────────────────────────────╮
│ ████████████████████░░░░  4 passed · 1 tools changed · 0 regressions │
╰───────────────────────────────────────────────────────────────────────╯
  ⚠ TOOLS_CHANGED  refund-flow  1 tool change(s)

Watching for changes...

Multi-Turn Testing

Most eval tools handle single-turn well. EvalView is built for multi-turn — clarification paths, follow-up handling, and tool use across conversations.

name: refund-needs-order-number
turns:
  - query: "I want a refund"
    expected:
      output:
        contains: ["order number"]
  - query: "Order 4812"
    expected:
      tools: ["lookup_order", "check_policy"]
      forbidden_tools: ["delete_order"]
      output:
        contains: ["refund", "processed"]
        not_contains: ["error"]
thresholds:
  min_score: 70

Each turn scored independently with conversation context. Per-turn judge scoring, not just final response.

Smart DX

EvalView doesn't just run tests — it understands your agent and configures itself.

Assertion Wizard — Tests From Real Traffic

Capture real interactions, get pre-configured tests. No YAML writing.

evalview capture --agent http://localhost:8000/invoke
# Use your agent normally, then Ctrl+C

Assertion Wizard — analyzing 8 captured interactions

  Agent type detected: multi-step
  Tools seen          search, extract, summarize
  Consistent sequence search -> extract -> summarize

  Suggested assertions:
    1. Lock tool sequence: search -> extract -> summarize  (recommended)
    2. Require tools: search, extract, summarize           (recommended)
    3. Max latency: 5000ms                                 (recommended)
    4. Minimum quality score: 70                           (recommended)

  Accept all recommended? [Y/n]: y
  Applied 4 assertions to 8 test files

Auto-Variant Discovery — Solve Non-Determinism

Non-deterministic agents take different valid paths. Let EvalView discover and save them:

evalview check --statistical 10 --auto-variant

  search-flow  mean: 82.3, std: 8.1, flakiness: low_variance
    1. search -> extract -> summarize  (7/10 runs, avg score: 85.2)
    2. search -> summarize             (3/10 runs, avg score: 78.1)

    Save as golden variant? [Y/n]: y
    Saved variant 'auto-v1': search -> summarize

Run N times. Cluster the paths. Save the valid ones. Tests stop being flaky — automatically.

Auto-Heal — Fix Flakes Without Leaving CI

Model got silently updated? Output drifted? --heal retries safe failures, proposes variants for borderline cases, and hard-escalates everything else. It also records when those retries were triggered by a likely model/runtime update.

evalview check --heal

  ⚠ Model update detected: gpt-5-2025-08-07 → gpt-5.1-2025-11-12 (3 tests affected)

  ✓ login-flow           PASSED
  ⚡ refund-request       HEALED   retried — non-deterministic drift
  ⚡ order-lookup         HEALED   retried — likely model/runtime update
  ◈ billing-dispute      PROPOSED saved candidate variant auto_heal_a1b2 (score 72)
  ⚠ search-flow          REVIEW   tool removed: web_search
  ✗ safety-check         BLOCKED  forbidden tool called — cannot heal

  3 resolved, 1 candidate variant saved, 1 needs review, 1 blocked.
  Model update: 2 of 3 affected tests healed via retry. Run `evalview snapshot` to rebase.
  Audit log: .evalview/healing/2026-03-25T14-30-00.json

Decision policy: Retry when tools match but output drifted (non-determinism or likely model/runtime update). Propose a variant when retry fails but score is acceptable. Never auto-resolve structural changes, forbidden tool violations, cost spikes, or score improvements. Full audit trail in .evalview/healing/.

Exit code: 0 only when every failure was resolved via retry. Proposed variants, reviews, and blocks always exit 1 — CI stays honest.

evalview check --budget 0.50

  $0.12 (24%) — search-flow
  $0.09 (18%) — refund-flow
  $0.31 (62%) — billing-dispute

  Budget circuit breaker tripped: $0.52 spent of $0.50 limit
  2 test(s) skipped to stay within budget

Supported Frameworks

Works with LangGraph, CrewAI, OpenAI, Claude, Mistral, HuggingFace, Ollama, MCP, and any HTTP API.

Framework details → | Flagship starter → | Starter examples →

How It Works

┌────────────┐      ┌──────────┐      ┌──────────────┐
│ Test Cases  │ ──→  │ EvalView │ ──→  │  Your Agent   │
│   (YAML)   │      │          │ ←──  │ local / cloud │
└────────────┘      └──────────┘      └──────────────┘

1. evalview init — detects your running agent, creates a starter test suite
2. evalview snapshot — runs tests, saves traces as baselines
3. evalview check — replays tests, diffs against baselines, emits the ship/don't-ship verdict, opens HTML report
4. evalview since — 2-second brief: what's changed since your last run (the daily habit anchor)
5. evalview watch — re-runs checks on every file save
6. evalview monitor — continuous checks in production with Slack alerts
7. evalview progress --since — diff any two points in history with a "worth a commit" gate
8. evalview drift — per-test sparklines, OLS slope, and incident markers
9. evalview slack-digest — post the daily verdict to your team channel

evalview snapshot list              # See all saved baselines
evalview snapshot show "my-test"    # Inspect a baseline
evalview snapshot delete "my-test"  # Remove a baseline
evalview snapshot --preview         # See what would change without saving
evalview snapshot --reset           # Clear all and start fresh
evalview replay                     # List tests, or: evalview replay "my-test"

Your data stays local by default. Nothing leaves your machine unless you opt in to cloud sync via evalview login.

Production Monitoring

evalview monitor                                         # Check every 5 min
evalview monitor --dashboard                             # Live terminal dashboard
evalview monitor --slack-webhook https://hooks.slack.com/services/...
evalview monitor --history monitor.jsonl                 # JSONL for dashboards
evalview monitor --incidents                             # Log confirmed regressions for `evalview autopr`

New regressions trigger Slack alerts. Recoveries send all-clear. No spam on persistent failures.

Every alert is a promise. The monitor requires two consecutive failing cycles before it pages a human — a single blip self-resolves silently and never interrupts anyone. If a test must alert on the first failure (auth, payments, PII, refund paths), mark it gate: strict in its YAML and it bypasses the gate, re-alerting every cycle until it passes.

Suppressed failures are never hidden: evalview slack-digest renders a Noise section listing every test the gate swallowed, how many times it self-resolved, and a visible false-positive rate (3 suppressed / 12 fired = 25% noise). See evalview/core/noise_tracker.py for the full design — confirmation gate, coordinated-incident collapse, and the .evalview/noise.jsonl metric.

Monitor config options →

Auto-PR Regression Tests From Production Incidents

evalview autopr closes the loop: production failure → pinned regression test → pull request, with zero LLM calls and zero manual YAML writing.

evalview monitor --incidents                             # Monitor writes .evalview/incidents.jsonl
evalview autopr --dry-run                                # Preview what would be generated
evalview autopr                                          # Write tests/regressions/*.yaml
evalview autopr --open-pr                                # + commit + push + gh pr create

The synthesizer is pure and deterministic — no network, no LLM — so it runs instantly in CI. For each confirmed regression in .evalview/incidents.jsonl it:

• builds a tests/regressions/<slug>.yaml that pins the query, the baseline tool sequence (expected.tools), the newly-appeared tools (forbidden_tools), and short phrases from the bad output (output.not_contains)
• tags the test suite_type: regression, gate: strict, and stamps the incident metadata into meta.incident so autopr can skip it on subsequent runs
• defaults to min_score: 90 — regression tests are a safety net, not a capability benchmark

# tests/regressions/refund-request-b3c4d5e6.yaml  (auto-generated)
name: regression_refund-request_2026-04-14
description: Auto-generated from production incident (REGRESSION) at 2026-04-14T12:34:56Z ...
input:
  query: "I want a refund for order #123"
expected:
  tools: [lookup_order, check_policy, process_refund]
  forbidden_tools: [escalate_to_human]   # appeared only in the failing trace
  output:
    not_contains: ["Sure, I've processed your refund for $999."]
thresholds:
  min_score: 90.0
suite_type: regression
gate: strict
tags: [incident, autopr]

Wire it into GitHub Actions: copy examples/github-workflow-autopr.yml to .github/workflows/evalview-autopr.yml — the workflow runs monitor then autopr --open-pr --require-new on a schedule. Every production regression arrives as a reviewable PR, and your hallucination test suite grows by itself.

The CLI is fully local. evalview monitor + evalview autopr run entirely on your machine — local files, gh pr create, no network calls, no cloud account required. The primitive is free and open-source forever.

Model Drift Detection

Closed models update silently. evalview model-check is a dedicated command that runs a fixed structural canary suite directly against the provider and tells you when the model itself has changed — no agent, no LLM judge, no calibration required.

# Save a baseline snapshot the day you deploy
evalview model-check --model claude-opus-4-5-20251101

# Run it weekly — detects any behavioral change against that baseline
evalview model-check --model claude-opus-4-5-20251101

Example output when drift is detected:

EvalView model-check
  Model:        claude-opus-4-5-20251101
  Provider:     anthropic
  Suite:        canary v1.public (15 prompts, sha256:6b8e925a5543…)
  Runs/prompt:  1
  Temperature:  0.0
  Cost:         $0.0228

vs reference (2026-04-10, 7d ago)
  Drift:      MODEL (MEDIUM confidence)
  Pass rate:  15/15 → 13/15 (-13.3%)
  Flipped:    tool_choice_refund_first_step, json_schema_order_summary

vs previous (2026-04-17, 0d ago)
  Drift:      NONE
  Pass rate:  13/15 → 13/15 (+0.0%)

How it works:

Every check runs at temperature=0 for determinism. Drift is classified as NONE / WEAK / MEDIUM / STRONG based on how many prompts flipped pass↔fail. No judge — the signal is structural, not probabilistic.

evalview model-check --model claude-opus-4-5-20251101 --dry-run      # Cost estimate before running
evalview model-check --model claude-opus-4-5-20251101 --pin           # Pin this run as the new reference
evalview model-check --model claude-opus-4-5-20251101 --reset-reference  # Start a fresh baseline
evalview model-check --model claude-opus-4-5-20251101 --json          # Machine-readable output for CI
evalview model-check --model claude-opus-4-5-20251101 --suite my-canary.yaml  # Bring your own suite

Ships with a bundled 15-prompt public canary covering tool selection, JSON schema, refusal behavior, and exact match. Add your own prompts with --suite. v1 supports Anthropic; OpenAI/Mistral/Cohere land in v1.1.

Full reference → docs/MODEL_CHECK.md

Simulation + Decision Rationale

Two features that close the gaps the April 2026 agent-eval reports kept calling out: pre-flight what-if testing and structured "why did the agent branch?" logging.

Pre-flight simulation — evalview simulate runs your test suite hermetically against declared mocks (tool calls, LLM responses, HTTP). Deterministic seed, zero cost, works in CI. Declare mocks in the test YAML, fan out with --variants N:

evalview simulate tests/ --variants 5 --seed 42

Decision rationale — every tool_choice / branch gets recorded with the chosen option, alternatives considered, and any model-reported reasoning (Anthropic thinking blocks auto-captured). Grouped across runs by input_hash so cloud analytics surfaces decision drift before your users notice. Local HTML replay shows it inline.

Supported adapters: Anthropic, OpenAI Assistants, LangGraph, CrewAI native.

docs/SIMULATE.md · docs/RATIONALE.md · examples/simulation/

Key Features

Python API

Use EvalView as a library — no CLI, no subprocess, no output parsing.

from evalview import gate, DiffStatus

result = gate(test_dir="tests/")

result.passed          # bool — True if no regressions
result.status          # DiffStatus.PASSED / REGRESSION / TOOLS_CHANGED
result.summary         # .total, .unchanged, .regressions, .tools_changed
result.diffs           # List[TestDiff] — per-test scores and tool diffs

result = gate(test_dir="tests/", quick=True)  # deterministic only, $0

result = await gate_async(test_dir="tests/")

from evalview.openclaw import gate_or_revert

make_code_change()
if not gate_or_revert("tests/", quick=True):
    # Change was reverted — try a different approach
    try_alternative()

OpenClaw Integration

Use EvalView as a regression gate in autonomous agent loops.

evalview openclaw install                    # Install gate skill into workspace
evalview openclaw check --path tests/        # Check and auto-revert on regression

from evalview.openclaw import gate_or_revert

make_code_change()
if not gate_or_revert("tests/", quick=True):
    try_alternative()  # Change was reverted

Pytest Plugin

def test_weather_regression(evalview_check):
    diff = evalview_check("weather-lookup")
    assert diff.overall_severity.value != "regression", diff.summary()

pip install evalview    # Plugin registers automatically
pytest                  # Runs alongside your existing tests

Model Comparison

Test the same task across multiple models with one parametrized test. No config files — just a model name and a query.

import pytest
import evalview

@pytest.mark.parametrize("model", ["claude-opus-4-6", "gpt-4o", "claude-sonnet-4-6"])
def test_my_task(model):
    result = evalview.run_eval(model, query="Summarize this contract in one sentence.")
    assert evalview.score(result) > 0.8

Provider is auto-detected from the model name. Requires ANTHROPIC_API_KEY / OPENAI_API_KEY depending on which models you use.

Score against expected output — token-overlap similarity, no LLM judge needed:

result = evalview.run_eval(
    "gpt-4o",
    query="What language is Python?",
    expected="Python is a high-level interpreted language.",
    threshold=0.4,
)

Custom scorer — assert specific behavior:

def has_json(output, expected):
    import json, re
    m = re.search(r"\{.*?\}", output, re.DOTALL)
    try: return 1.0 if json.loads(m.group()) else 0.0
    except: return 0.0

result = evalview.run_eval("claude-opus-4-6", query="Return JSON: {name, age}", scorer=has_json)
assert evalview.score(result) == 1.0

Run all models in parallel and compare:

results = evalview.compare_models(
    query="Explain quantum entanglement in one sentence.",
    models=["claude-opus-4-6", "gpt-4o", "claude-sonnet-4-6"],
)
evalview.print_comparison_table(results)   # Rich table: score, latency, cost
best = results[0]                          # sorted best-first

┏━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━┓
┃ Model              ┃ Score ┃  Latency ┃      Cost ┃ Pass? ┃
┡━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━┩
│ claude-opus-4-6    │  1.00 │    842ms │ $0.00312  │   ✓   │
│ gpt-4o             │  1.00 │    631ms │ $0.00087  │   ✓   │
│ claude-sonnet-4-6  │  1.00 │    514ms │ $0.00063  │   ✓   │
└────────────────────┴───────┴──────────┴───────────┴───────┘

ModelResult fields: .model, .output, .score, .latency_ms, .cost_usd, .passed, .error

Full example →

Claude Code (MCP)

claude mcp add --transport stdio evalview -- evalview mcp serve

8 tools: create_test, run_snapshot, run_check, list_tests, validate_skill, generate_skill_tests, run_skill_test, generate_visual_report

# 1. Install
pip install evalview

# 2. Connect to Claude Code
claude mcp add --transport stdio evalview -- evalview mcp serve

# 3. Make Claude Code proactive
cp CLAUDE.md.example CLAUDE.md

Agent-Friendly Docs

Works with your coding agent out of the box. Ask Cursor, Claude Code, or Copilot to add regression tests, build a new adapter, or debug a failing check — EvalView ships the architecture maps and task recipes they need to get it right on the first try.

• AGENT_INSTRUCTIONS.md — architecture map, contracts, invariants, verification commands
• Agent Recipes — task-specific playbooks for common extensions
• Add an Adapter
• Add an Evaluator
• Debug Check vs Snapshot Mismatch
• Extend the HTML Report
• Integrate Ollama

Documentation

Contributing

• Bug or feature request? Run evalview feedback or open an issue
• Questions? GitHub Discussions
• Setup help? Email hidai@evalview.com
• Contributing? See CONTRIBUTING.md

License: Apache 2.0

Star History