Free & open source/MIT/Works in 30+ agents

Your AI metrics are lying.
Ship ones that aren't.

Actuals is a free plugin for Claude Code, Codex, Cursor, and 30+ other coding agents: it interviews you about your business, writes a versioned measurement spec, audits your dashboards against 20 named vanity-metric anti-patterns — and generates the instrumentation to measure what's real.

Get the plugin — free → Star on GitHub ★ See an audit in action
The measurement gap, in three receipts
R-01 · METR
19% slower

developers using AI in METR's randomized trial — while believing they were 20% faster

R-02 · MICROSOFT
$72 / hour

Microsoft's flat multiplier behind Copilot's "assisted value" dollar figure

R-03 · MCKINSEY
39%

of companies can attribute any profit impact to AI — McKinsey, State of AI

Works with
PostHogLangfuseBraintrustArize PhoenixStripeGitHubLinearJiraSupabasedbtSnowflakeBigQueryAmplitudeMixpanel+ more
§ 01
The loop

Five commands. One honest loop.

Each skill triggers on plain English too — say "are these numbers real?" and the audit just runs.

01  /actuals:design

A structured interview — what the AI does, who benefits, and the decision your metrics must inform — becomes a versioned measurement spec with formulas, owners, and guardrails.

02  /actuals:audit

Point it at dashboards, tracking plans, SQL, or the ROI deck. Every finding cites a named anti-pattern, quotes your artifact, carries a published receipt, and ships a fix.

03  /actuals:instrument

Events with typed constants, SQL/dbt models implementing each formula verbatim, a pre-launch baseline snapshot, and an eval harness with judge–human calibration built in.

04  /actuals:connect

Vetted MCP configs for the tools you already run — merged into your project non-destructively. Secrets stay as env-var placeholders you export yourself.

05  /actuals:scorecard

A self-contained dashboard.html — metrics vs baselines and targets, tripwires, open findings, the claims ledger. Renders recorded values only, and says so.

↻  then: re-audit monthly

The audit's review mode checks spec drift, definition rot, stale owners, and overdue calibrations — the maintenance loop that keeps the spec from becoming the thing it replaced.

§ 02
Start anywhere

From zero — or from the dashboard you already have

Starting fresh

Design the spec

One interview produces metrics/MEASUREMENT.md — the single source of truth your reporting has been missing.

  • +The decision each metric informs — or the metric gets deleted
  • +3–5 outcome metrics with formulas and honest denominators
  • +A guardrail with a tripwire for every headline number
  • +A claims ledger recording what you can't say without a baseline
Install & run design →
Already shipping

Audit what you have

audit findingcritical
VM-02 · Minutes-times-Wage Dollarization
where:  dashboard-export.csv:4
quote:  "1,240 hrs × $72/hr = $89,280 / month"
receipt:[MSFT-COPILOT] flat-multiplier value math
fix:    cost per resolved ticket — constants live
        in the spec, re-measured quarterly
  • +20 named anti-patterns with stable IDs — cite them like CVEs
  • +Published evidence behind every finding
  • +A concrete replacement for every number it kills
Browse the catalog
§ 03
Before / after

What an audit does to a dashboard

From the demo fixture that ships with the plugin — a SaaS with an AI support assistant and a very confident dashboard.

What the dashboard said
What Actuals replaces it with
Hours saved per week (agent survey): 1,240 hrsVM-01
Median handle time: 14.2 → 9.8 min — system-measured, 8-week window
Estimated value: $89,280 / month (× $72/hr)VM-02
Cost per resolved ticket: $8.10 → $6.35 — constants re-measured
Total conversations since launch: 152,340VM-06·08
Deflection-without-reopen: 34% of eligible tickets — weekly, honest denominator
CSAT 4.7/5 among active users (excl. churned)VM-10
CSAT 4.1 across all offered tickets — 18% disabled the assistant, now tracked
Run it yourself from a clone: /actuals:audit examples/acme-support-ai/ — answer key included.
§ 04
The catalog

20 named ways AI metrics lie

Stable IDs, detection signals, severity, published evidence, and a concrete fix for each.

VM-01
CRITICAL

Self-Reported Time Savings

"Devs report saving 11 hrs/week." Perception inverts reality — METR's trial participants were 19% slower while feeling 20% faster.

→ measure handle time from system data, or drop the claim
VM-02
CRITICAL

Minutes-times-Wage Dollarization

Actions × minutes × $/hour. Every constant is an assumption in a dollar costume — and it can break silently for months.

→ dollarize only through real unit costs with re-measured constants
VM-06
WARNING

Denominator-Free Count

"10,000 AI conversations!" Out of how many tickets? A numerator without a denominator is a press release, not a metric.

→ share of eligible volume, fixed window, cohort trend
§ 05
Integrations

Nothing bundled that phones home

Data in

Connect what you run

A curated, verified MCP registry — written into your config on your confirmation. Env-var placeholders only, never literal keys.

16 verified vendorsnon-destructive merge
Definitions out

Your BI does the computing

Instrument publishes metric definitions as PostHog insights and dashboards — your data carries the live numbers, the spec stays the source of truth.

PostHog dashboardsSQL / dbt models
Checks as tools

Ships with its own MCP server

Run the deterministic checks from CI or any agent — no skills required, zero dependencies, stdio.

spec_lintvanity_scan
§ 06
On honesty

Practicing what it preaches

Actuals ships with its own measurement spec, written with its own method. GitHub stars and installs are listed there under "Vanity Metrics We Will Not Use." Its claims ledger commits to never claiming it improved your metrics without your baselines. There's no telemetry in the plugin — and this page loads no external resources and runs no analytics.

§ 07
Get started

Get Actuals in your agent

Same five skills everywhere — Actuals follows the open Agent Skills standard, so it installs into whichever agent you code with.

claude
> /plugin marketplace add Hamza-Saraswat/actuals
> /plugin install actuals@actuals-marketplace

Then just ask: "what should we measure for our new AI feature?" — or hand it your dashboard and ask if the numbers are real.

Star on GitHub ★