Epoch -- Time Estimation MCP Server

Epoch is a time estimation MCP server that helps engineering leads, agents, and planners who need calibrated duration estimates produce PERT/reference-class estimates and record actuals.

TL;DR: Epoch — time estimation MCP server. Best for engineering leads, agents, and planners who need calibrated duration estimates.

Epoch helps AI agents understand time.

AI can write code, analyze data, and build apps -- but it can't tell you how long something will take, what it'll cost, or whether a deadline is realistic. Epoch fixes that.

Show Me

You: "How long will a 15,000 line project take with a small team?"

Claude (using Epoch):
  COCOMO II nominal: 100 person-months (traditional team)
  AI-adjusted: 9 person-months (LLM-assisted workflow)
  Schedule risk: medium (MAPE 15%, based on 28 historical features)
  Cost estimate (Claude Sonnet): ~$340 in API tokens

You: "How many working days between now and launch?"

Claude (using Epoch):
  Start: May 1, 2026
  End: July 15, 2026
  Business days: 54
  Weekend days: 16
  Holidays: 1 (Independence Day)
  Total calendar days: 75

You: "Which AI model should we use for a 50k-token job?"

Claude (using Epoch):
  gemini-2.0-flash:  $0.02 -- fast, budget-friendly
  gpt-4o-mini:       $0.02 -- fast, budget-friendly
  gemini-2.5-pro:    $0.37 -- standard, high quality
  claude-sonnet-5:    $0.57 -- standard, high quality
  Recommendation: gemini-2.0-flash for cost, claude-sonnet-5 for quality

Why Epoch?

Every AI agent hallucinates timelines. "This should take about 2 hours" becomes 2 days. Epoch gives AI grounded, data-driven estimates instead of guesses. It packages established estimation methods (PERT, COCOMO II, Monte Carlo, reference class forecasting) into 26 tools any AI can call -- so your assistant stops guessing and starts calculating.

Works out of the box. Epoch ships with a bundled reference database: task-type correction factors and scope baselines from the maintainer's recorded estimate-vs-actual history plus curated public research, and tool-usage/latency telemetry samples (117,791 across the bundled canary-test and session benchmarks) that keep the token-time bridge grounded. You get reasonable estimates from day one — no data collection or account setup required. If you choose to record your actuals, Epoch's self-improvement engine learns your patterns and gets even more precise over time.

What is MCP?

MCP (Model Context Protocol) is how AI assistants like Claude connect to external tools. Think of it like a plugin system -- you add Epoch with one command, and suddenly your AI assistant can estimate timelines, calculate business days, compare model costs, and predict whether your project will finish on time.

Quick Start

30-second setup -- works in Claude Code, Cursor, VS Code, and Windsurf:

claude mcp add epoch -- npx @kyanitelabs/epoch

That's it. Your AI assistant now has 26 time estimation tools.

Or add it to your project's .mcp.json:

{
  "mcpServers": {
    "epoch": {
      "command": "npx",
      "args": ["@kyanitelabs/epoch"]
    }
  }
}

Agent Skill

Epoch also ships a public agent skill at skills/epoch/SKILL.md. Use $epoch in compatible agent hosts when you want the agent to choose the right Epoch MCP or CLI workflow for time estimates, business-day math, model-cost comparison, schedule risk, and estimate-vs-actual feedback.

What Can Epoch Do?

Technical Reference

Everything below is for developers who want to understand the internals, use the CLI or REST API, or contribute to Epoch.

Architecture

Six-layer design with 26 tools for time estimation, scheduling, cost analysis, and feedback:

Tool Reference

Layer 1 -- Core Temporal

get_current_time -- Current wall-clock time in any IANA timezone

Input:  { timezone: "America/New_York" }
Output: {
  iso: "2026-05-01T08:30:00.000-04:00",
  humanReadable: "Fri, May 1, 2026, 8:30 AM EDT",
  timezone: "America/New_York",
  utcOffset: "-04:00"
}

convert_timezone -- Convert a timestamp between IANA timezones

Input:  { timestamp: "2026-05-01T12:00:00Z", target_tz: "Asia/Tokyo" }
Output: {
  iso: "2026-05-01T21:00:00.000+09:00",
  timezone: "Asia/Tokyo",
  utcOffset: "+09:00",
  humanReadable: "Fri, May 1, 2026, 9:00 PM JST"
}

parse_duration -- Parse human-readable duration strings

Input:  { duration_string: "2h30m" }
Output: {
  input: "2h30m",
  totalSeconds: 9000,
  humanReadable: "2 hours 30 minutes"
}

time_math -- Date arithmetic operations

Input:  { operation: "add_days", date: "2026-05-01", value: 7 }
Output: {
  result: "2026-05-08T00:00:00.000Z",
  operation: "add_days",
  input: "2026-05-01"
}

Supported operations: add_days, add_business_days, diff, convert_tz, parse_nl, format_duration

Layer 2 -- Calendar Math

add_business_days -- Add N business days with holiday awareness (US, UK, FR, DE, JP)

Input:  { start_date: "2026-05-01", days: 5, country: "US" }
Output: {
  startDate: "2026-05-01",
  endDate: "2026-05-08",
  businessDays: 5,
  countryCode: "US",
  holidaySupport: "holiday_calendar",
  humanReadable: "5 business days from 2026-05-01 to 2026-05-08 (US)."
}

count_business_days -- Count business days between two dates

Input:  { start_date: "2026-05-01", end_date: "2026-05-15", country: "US" }
Output: {
  startDate: "2026-05-01",
  endDate: "2026-05-15",
  businessDays: 10,
  countryCode: "US",
  holidaySupport: "holiday_calendar",
  humanReadable: "10 business days between 2026-05-01 and 2026-05-15 (US)."
}

Both tools always include holidaySupport ("holiday_calendar" when the country has a bundled holiday set — US/UK/FR/DE/JP — otherwise "weekends_only") in successful output; country must match ^[A-Za-z]{2}$ — "USA"-style codes are rejected with a readable error.

Layer 3 -- Estimation

pert_estimate -- PERT three-point estimation with confidence intervals and urgency scoring

Input:  {
  optimistic: 2,
  most_likely: 4,
  pessimistic: 12,
  unit: "hours"
}
Output: {
  expected: 5,
  variance: 2.78,
  stdDeviation: 1.67,
  confidence95: [1.67, 8.33],
  confidence99: [0, 10],
  unit: "hours",
  urgencyCategory: "medium",
  riskLevel: "high",
  humanReadable: "Expected 2.86–7.14 hours (80% confidence interval); point estimate 5 hours (ledger-recorded basis; adjustedEstimate 5 applies the correction factor). No task_type was supplied, so this interval is derived from the PERT variance (optimistic/most_likely/pessimistic spread) instead of empirical data.",
  developerProfile: { mode: "ai_native", correctionFactor: 1 },
  adjustedEstimate: 5,
  rawEstimate: 5,
  correctionFactor: 1,
  n: 0,
  interval: { p50: { lower: 3.87, upper: 6.13 }, p80: { lower: 2.86, upper: 7.14 }, p90: { lower: 2.25, upper: 7.75 }, source: "pert_variance" },
  intervalNote: "No task_type was supplied, so this interval is derived from the PERT variance (optimistic/most_likely/pessimistic spread) instead of empirical data.",
  basisNote: "Interval and point estimate are on the ledger-recorded basis (raw PERT expected × unit factor). adjustedEstimate (5 hours) additionally applies the correction factor (1) and is display-only — it is never recorded or calibrated against."
}

(Example from a fresh install with no accumulated feedback; with ≥5 exclusion-filtered matched pairs for the task type, interval.source becomes "empirical_ratio_quantile", n reports the sample size, and intervalPopulation names the ratio population used.)

cocomo_estimate -- COCOMO II software sizing with LLM-adapted cost drivers

Input:  {
  kloc: 15,
  reasoning_complexity: 1.2,
  context_completeness: 1.0,
  transformation_impact: 0.8,
  iterative_cycles: 1.5,
  human_oversight: 1.2
}
Output: {
  kloc: 15,
  personMonthsNominal: 99.9,
  personMonthsLlmAdjusted: 8.9,
  effortMultipliers: {
    reasoning_complexity: 1.2,
    context_completeness: 1.0,
    transformation_impact: 0.8,
    iterative_cycles: 1.5,
    human_oversight: 1.2,
    product: 1.728
  },
  developerProfile: { mode: "ai_native", correctionFactor: 1.45 }
}

LLM-adapted cost drivers include reasoning complexity, context completeness, transformation impact, iterative cycles, and human oversight requirements.

sprint_forecast -- Sprint velocity forecasting from historical data

Input:  {
  backlog_points: 100,
  velocity_history: [20, 25, 22, 23],
  sprint_length_days: 14,
  hours_per_sprint: 80
}
Output: {
  backlogPoints: 100,
  averageVelocity: 22.5,
  requiredSprints: 4.4,
  pessimisticSprints: 4.9,
  hoursPerPoint: 3.56,
  totalHours: 355.6,
  completionDays: 62,
  sprintLengthDays: 14,
  developerProfile: { mode: "ai_native", sprintVelocityPoints: 80, correctionFactor: 1.45 }
}

critical_path -- Critical Path Method with merge-bias adjustment for parallel tasks

Input:  {
  tasks: [
    { name: "A", duration: 5, predecessors: [] },
    { name: "B", duration: 3, predecessors: ["A"] },
    { name: "C", duration: 4, predecessors: ["A"] }
  ]
}
Output: {
  critical_path: ["A", "C"],
  total_duration: 9,
  slack_per_task: { A: 0, B: 1, C: 0 },
  merge_bias_adjustment: 0
}

monte_carlo_schedule -- Monte Carlo simulation with seeded PRNG for deterministic, reproducible results

Input:  {
  tasks: [
    { name: "A", optimistic: 2, most_likely: 4, pessimistic: 8 },
    { name: "B", optimistic: 1, most_likely: 3, pessimistic: 6 }
  ],
  iterations: 10000
}
Output: {
  p10: "5.9",
  p50: "7.91",
  p80: "9.39",
  p95: "10.75",
  riskEvents: [
    { description: "Task \"A\" exceeded 1.5x PERT expected in 10% of simulations", probability: 0.1, impactDays: 0.05 },
    { description: "Task \"B\" exceeded 1.5x PERT expected in 10% of simulations", probability: 0.1, impactDays: 0.04 }
  ],
  criticalPathProbability: null
}

(criticalPathProbability is null unless a target_hours deadline is supplied — then it is the real P(total ≤ target); riskEvents[].impactDays is per-task expected overrun, sorted by impact.)

Layer 4 -- Analytics

reference_class_estimate -- Reference class forecasting with planning fallacy correction

Input:  {
  task_type: "feature",
  complexity: 3
}
Output: {
  rawEstimate: 2,
  correctedEstimate: 2,
  correctionFactor: 1,
  sampleSize: 0,
  baselineSource: "inferred_scope_medium_real_tasks",
  scopeUsed: "medium",
  scopeInferred: true,
  confidence: "pessimistic",
  developerProfile: { mode: "ai_native", estimationMape: 15, underestimationBias: 0.2, correctionFactor: 1 },
  adjustedEstimate: 2,
  basisNote: "correctedEstimate (2 hours) is the ledger-recorded and displayed basis (rawEstimate × correctionFactor). adjustedEstimate (2 hours) additionally applies the developerProfile factor (1) and is display-only — it is never recorded or calibrated against.",
  intervalNote: "Fewer than 5 exclusion-filtered historical \"feature\" reference_class_estimate pairs are available yet, so no empirical confidence interval could be computed."
}

(Fresh-install output; with accumulated feedback the correction factor and sample size come from your own matched pairs and the empirical interval is populated — correctedEstimate is always the recorded basis.)

Valid task_type values: feature, bugfix, refactor, migration, infrastructure, documentation, testing, design.

estimate_from_context -- Classify a free-text task description and delegate to reference class estimation

Input:  {
  context: "Add OAuth2 login support to the API, including refresh token rotation and a new /auth/callback endpoint"
}
Output: {
  tool: "estimate_from_context",
  rawEstimate: 2,
  correctedEstimate: 2,
  correctionFactor: 1,
  sampleSize: 0,
  baselineSource: "inferred_scope_medium_real_tasks",
  scopeUsed: "medium",
  scopeGuide: "For feature tasks: small=~2.3h, medium=~6h, large=~10.6h, xl=~17h",
  classification: {
    classified_task_type: "feature",
    classified_complexity: 3,
    confidence: "medium",
    signals: ["task_type_matched:feature"],
    task_type_from_hint: false,
    complexity_from_hint: false
  },
  note: "Using reference database correction factors. Submit actuals via record_actual to improve accuracy."
}

Classifies task_type and complexity from free text (an issue body, PR/diff description, or task summary) using a local, deterministic keyword/signal heuristic -- no LLM call is made. Caller-supplied task_type/complexity hints always override the classification. The resolved inputs are then delegated to the same reference-class-forecasting path used by reference_class_estimate, so the response carries the same estimate fields plus a classification provenance block explaining how the tool read the context. When classification confidence is low, an additional lowConfidenceNote field is returned rather than silently guessing.

calibrate_estimates -- Team-specific accuracy calibration from historical estimated vs actual data

Input:  {
  task_type: "feature",
  team_id: "backend"
}
Output: {
  correctionFactor: 1.45,
  accuracyTrend: "stable",
  velocityTrend: "stable",
  recommendations: [
    "Using reference database correction factor (1.45x) — 8 matched samples, need 5.",
    "Record actuals via POST /v1/feedback/record-actual to refine for your team's patterns."
  ]
}

token_time_bridge -- Map LLM token budgets to wall-clock time for 16 model families

Input:  {
  tokens: 50000,
  model: "claude-sonnet-4-20250514",
  tool_calls: 10,
  reasoning_depth: "deep"
}
Output: {
  estimatedSeconds: 697,
  estimatedMinutes: 11.6,
  confidence: "likely",
  urgency: "short",
  breakdown: {
    promptTokens: 15000,
    completionTokens: 35000,
    toolOverheadSeconds: 2
  }
}

Layer 5 -- Cost & Risk

token_cost_estimate -- Token cost estimation for LLM API calls

Input:  {
  tokens: 50000,
  model: "claude-sonnet-5"
}
Output: {
  tokens: 50000,
  model: "claude-sonnet-5",
  estimatedSeconds: 695,
  estimatedMinutes: 11.6,
  estimatedCost: 0.57,
  costBreakdown: { inputCost: 0.045, outputCost: 0.525, toolCallOverheadCost: 0 },
  confidence: "likely"
}

compare_models -- Side-by-side cost and capability comparison across LLM models

Input:  {
  tokens: 50000,
  sort_by: "cost"
}
Output: {
  tokens: 50000,
  models: [
    { model: "gemini-2.0-flash", estimatedCost: 0.0155, qualityTier: "fast", tokensPerSecond: 230 },
    { model: "deepseek-v3", estimatedCost: 0.0189, qualityTier: "standard", tokensPerSecond: 97 },
    { model: "gpt-4o-mini", estimatedCost: 0.0233, qualityTier: "fast", tokensPerSecond: 180 }
  ],
  sortBy: "cost"
}

accuracy_trend -- Track estimation accuracy over time from recorded feedback data

Input:  { team_id: "backend", window_size: 50 }
Output: {
  overallTrend: "improving",
  currentMape: 26.5,
  industryBaselineMape: 25,
  totalEstimates: 1049,
  totalWithActuals: 1049,
  windows: [{ period: "Window 1 (estimates 1-50)", mape: 32, bias: 5.3, sampleSize: 50 }]
}

schedule_risk -- Schedule risk scoring for project timelines

Input:  {
  estimated_hours: 40,
  task_type: "feature"
}
Output: {
  estimatedHours: 40,
  riskLevel: "low",
  confidenceIntervals: { p50: 40, p80: 45.1, p95: 49.9 },
  historicalAccuracy: { mape: 15, sampleSize: 42 },
  recommendation: "Low risk. Estimate is within normal variance.",
  humanReadable: "Schedule risk: low. MAPE: 15% (based on 42 historical records). Confidence intervals: p50=40h, p80=45.1h, p95=49.9h."
}

cocomo_validate -- Validate COCOMO II estimates against reference data

Input:  {}
Output: {
  projectsEvaluated: 182,
  mape: 85.55,
  bias: 53.5,
  byProjectType: {
    organic: { mape: 86.57, count: 22 },
    semidetached: { mape: 84.75, count: 106 },
    embedded: { mape: 86.71, count: 54 }
  },
  recommendedAdjustments: []
}

cocomo_ground_truth -- Benchmark all COCOMO variants (Basic, COCOMO II nominal, AI 12x speedup, AI + developer-profile gradients) against the same real historical projects, with per-dataset and per-type breakdowns

Input:  {}
Output: {
  projectsEvaluated: 182,
  models: [
    { name: "COCOMO Basic", mape: 85.55, mmre: 0.856, pred25: 0.313, pred50: 0.544, bias: 53.5, count: 182 },
    ...
  ],
  byDataset: { ... },
  byType: { ... },
  winner: "AI + Profile (human)",
  conclusion: "Best model: AI + Profile (human) (MAPE=79.66%). ...",
  humanReadable: "..."
}

ai_native Mode

Epoch tools support dual estimation modes to account for the fundamentally different velocity of AI-assisted vs human-only development.

When ai_native=true (default), tools use Epoch's reference database with tool-aware correction factors. These baselines reflect AI agent workflows: faster iteration, higher output volume, and different error profiles.

When ai_native=false, tools apply human developer baselines:

Tools that support ai_native: pert_estimate, cocomo_estimate, sprint_forecast, reference_class_estimate, schedule_risk.

Hybrid workflows: ai_native accepts a float from 0.0 (fully human) to 1.0 (fully AI-native). Values like 0.5 produce interpolated profiles for mixed AI/human workflows. Boolean values (true/false) remain supported for backward compatibility.

Self-Improvement Engine

Epoch learns your patterns the more you use it. The bundled reference database combines a small set of maintainer estimate-vs-actual pairs (the seed of the correction factors), curated public research baselines, and 117,791 tool-usage/latency telemetry samples (canary-test and session benchmarks — these ground latency and token-time figures, not estimation accuracy). It gives you a sane starting point on day one — and it gets sharper as you record actuals.

If you record your actuals, Epoch personalizes further:

1. Estimate -- Generate an initial estimate with any estimation tool
2. Record -- Track the actual outcome (record_actual)
3. Learn -- Self-improvement computes personalized correction factors from your data
4. Improve -- Future estimates apply your team's actual patterns
5. Trend -- accuracy_trend tracks whether your accuracy is improving over time

Your estimates + your actuals -> Your correction factors -> Better estimates -> Repeat

The loop can close itself. Recording actuals is the step everyone forgets, so Epoch can do it for you: epoch auto-actuals --session <id> records wall-clock-derived actuals for a session's unfinished estimates (agent hosts can wire it into a session-end hook). Auto-recorded actuals are sanity-bounded (0.05–12h, <10x the estimate), provenance-labeled auto_wallclock, never overwrite a real actual, and feedback_health reports them separately (byProvenance) so automated data can't silently skew your calibration.

If the estimate writer cannot stamp inputs.session_id, pass the session's real time span explicitly: epoch auto-actuals --session <id> --window-start <ts> --window-end <ts>. Each <ts> may be epoch milliseconds or an ISO timestamp, and both flags must be supplied together with the end no earlier than the start. Epoch uses this window only when the durable ledger contains no estimate with the exact session ID, including a finalized or expired estimate, and only for pending estimates with no session stamp; estimates stamped to another session are never selected. Window-selected candidates still pass the same sanity and duplicate guards as exact session matches.

Your git history is another actuals source. epoch mine-git --repo <path> --since <date> walks a local git repository — offline and read-only by construction — and joins pending estimates to closed work (merged PRs and plain merges) via the identity keys planning flows already stamp: task_label, branch, or issue_ref. Two cycle-time windows are derived per unit and reported separately, never blended: first-commit→merge (provenance git_derived) and open→merge (provenance git_derived_review_inclusive, computed wherever a branch-creation anchor survives locally — plain clones have none, and the window is never fabricated). Pick the recorded window with --window dev|review-inclusive; the other rides along in the actual's notes. Mined actuals are sanity-bounded ([0.05h, 720h] — calendar windows legitimately span nights and weekends — and <10x the estimate), never overwrite a real actual, and feedback_health segments them in its own byProvenance bucket (gitDerived) so calendar-window noise cannot silently skew calibration.

Estimates lead with honest ranges. When at least 5 matched pairs exist for a task type, pert_estimate and reference_class_estimate open with a calibrated 80% interval ("Expected 1.6–4.2 hours (80% confidence interval); point estimate 2.5 hours") derived from your own historical estimate-vs-actual ratios — and say plainly when there isn't enough data yet.

The engine detects systematic biases (chronic under-estimation, accuracy degradation) and surfaces actionable recommendations.

You do not need to share data with anyone for this to work. Self-improvement runs entirely locally using your own ~/.epoch/ data.

The correction loop, measured

The self-improvement claim above isn't marketing copy -- it's backed by a runnable receipt. scripts/backtest-pert-correction.mjs makes a read-only temp copy of your ~/.epoch ledger, chronologically splits matched pert_estimate (estimate, actual) pairs 80/20, trains the learned per-(tool, task_type) correction factor on the training split only, and reports MdAPE on the held-out test split it never trained on:

npx tsx scripts/backtest-pert-correction.mjs

Measured on the maintainers' production ledger (697 held-out matched pairs at time of writing): MdAPE improved from 105.2% (uncorrected) to 80.5% (learned correction) on data the correction factor never saw during training. This is the mechanism EPOCH_PERT_LEARNED_CORRECTION gates behind before it's recommended on by default -- the script also checks that the corrected median actual/predicted ratio lands in [0.7, 1.3], and reports HOLD (not recommended yet) when that second guard hasn't cleared, so the flag doesn't ship as "on" until both hold. Run the script against your own ledger for your own numbers; they move as more actuals get recorded, which is the point.

reference_class_estimate's correction factors are the same learned mechanism applied to a different tool. Track its current calibration with epoch data status or feedback_health (per-tool MAPE/MdAPE, bias, and trend), or generate a full calibration decision-surface report with node scripts/build-calibration-dashboard.mjs -- also strictly read-only against your ledger.

Data Pipeline

Epoch uses a three-layer data strategy so it's accurate from the start and gets better over time:

1. Bundled reference database (works immediately, no setup): Epoch ships with a pre-built reference database: correction factors seeded from the maintainer's own estimate-vs-actual history, scope baselines from curated public research, and 117,791 tool-usage/latency telemetry samples (canary tests and session benchmarks across 8 task types and 5 complexity levels). Telemetry sample counts are reported separately from matched estimate-actual pair counts everywhere Epoch shows a data size.

2. Local self-improvement (automatic, private): As you use Epoch and record actuals, the self-improvement engine recalibrates correction factors from your data. This runs entirely locally in ~/.epoch/ — nothing leaves your machine. The engine triggers automatically every 100 tool calls or 24 hours.

• Auto-recording: Use scripts/auto-record-actual.mjs to automatically record actual time against pending estimates.
• Source tagging: Set EPOCH_SOURCE=<project-name> to tag estimates by project.
• Inspect your data: epoch data where and epoch data status show what's stored locally.

3. Community contributions (optional, opt-in): You can optionally share anonymized data to help improve baselines for all users. Community data is stripped of all identifying information — only task type, complexity, estimated hours, actual hours, and date remain. See CONTRIBUTING-data.md for format and privacy requirements.

epoch share-data --validate --description "My anonymized estimation data"

This is completely optional. Epoch works great without it.

Surfaces

Epoch exposes the same 26 tools through three interfaces:

Default behavior: running epoch with no arguments starts the MCP stdio server.

CLI

# PERT estimate
epoch pert-estimate --optimistic 2 --most-likely 4 --pessimistic 12 --unit hours

# Token-to-time bridge
epoch token-time-bridge --tokens 50000 --model claude-sonnet-4-20250514

# Monte Carlo simulation
epoch monte-carlo-schedule --tasks '[{"name":"A","optimistic":2,"most_likely":4,"pessimistic":8}]'

# COCOMO II estimate
epoch cocomo-estimate --kloc 15

# Schedule risk score
epoch schedule-risk --estimated-hours 40

# List all tools
epoch list-tools

# Pretty table output
epoch pert-estimate --optimistic 2 --most-likely 4 --pessimistic 12 --pretty

REST API

# Start the server
epoch serve --port 3099
# or: EPOCH_TRANSPORT=http EPOCH_PORT=3099 epoch

# Call any tool
curl -X POST http://localhost:3099/v1/tools/pert_estimate \
  -H "Content-Type: application/json" \
  -d '{"optimistic": 2, "most_likely": 4, "pessimistic": 12, "unit": "hours"}'

# Health check
curl http://localhost:3099/health

# OpenAPI spec
curl http://localhost:3099/openapi.json

Agent-First

Epoch is built for agents as first-class callers, not humans typing in a terminal as an afterthought.

Why agents need time-sense. An LLM has no grounded sense of duration or cost -- it will say "quick fix" for a two-day migration and "big project" for a two-hour config change with equal confidence, because it has no feedback loop telling it otherwise. That's fine for a chat answer; it breaks down the moment an agent is planning multi-step work, sequencing a sprint, or deciding whether a deadline is realistic. Epoch gives the agent a calculator instead of a guess: PERT/COCOMO/Monte Carlo math, a reference-class baseline built from real task data, and a feedback loop that corrects itself as the agent (or its operator) records actuals.

EPOCH_TELEMETRY=1 for headless/agent operators. Telemetry is off by default and requires informed consent. For a human at a terminal, that consent is epoch telemetry enable, which shows the data and asks for confirmation. An agent should never be the one clicking "yes" to that prompt on its own behalf -- there is deliberately no MCP tool that enables telemetry, so an agent cannot self-consent. For headless or agent-operated deployments, the operator opts in out-of-band by setting EPOCH_TELEMETRY=1 in the server's environment (for example, the env block of the MCP server config) before the agent ever starts. Consent stays with the human who configures the deployment, not the agent that runs inside it.

MCP client qualification. Epoch's telemetry schema (v2) records client_name/client_version from the MCP clientInfo your host reports at connection time, plus transport (stdio/http). This is agent qualification, not agent identification: it lets aggregate accuracy stats count "5.7h median across N agent-driven feature estimates" as first-class agent data rather than lumping it in with anonymous CLI usage, without adding any new per-user identifying signal. MCP clients that report clientInfo (Claude Code, Cursor, and most current hosts do) get this for free; clients that don't are still fully functional, they just show up as client_name: null.

Epoch also provides built-in discoverability endpoints so agents can find and use the HTTP API without prior configuration:

Installation

git clone https://github.com/KyaniteLabs/Epoch.git
cd Epoch
pnpm install
pnpm run build

Development

pnpm test          # Run the Vitest suite
pnpm run build     # Build with tsup
pnpm run typecheck # TypeScript strict mode check
pnpm run dev       # Run development server
pnpm run inspector # Open MCP Inspector for interactive testing

Tech Stack

• Runtime: Node.js 22+ (ESM; engines.node >=22 — Node 20 reached EOL April 2026)
• Language: TypeScript 6 (strict mode, noUncheckedIndexedAccess, verbatimModuleSyntax)
• Validation: Zod 4 with .describe() on every field
• MCP SDK: @modelcontextprotocol/sdk 1.12+
• HTTP: Hono (lightweight, multi-runtime)
• CLI: Commander.js
• Date Handling: date-fns 4.x + date-fns-tz 3.x
• Build: tsup (ESM output)
• Testing: vitest 4.x with v8 coverage (97% statements, 88% branches)

Configuration

Telemetry & Privacy

Epoch can share anonymized estimate/actual pairs to improve accuracy for all users. This is off by default and requires explicit opt-in.

Agent-operator consent model: there is deliberately no MCP tool that enables telemetry -- an agent must not be able to self-consent on a human's behalf. Humans opt in interactively with epoch telemetry enable. Agent/headless operators opt in out-of-band by setting EPOCH_TELEMETRY=1 in the server's environment before the agent starts (see Agent-First). Either way, consent belongs to the person who configures the deployment.

epoch telemetry enable     # Opt in (shows exactly what will be shared)
epoch telemetry preview    # Preview anonymized data before enabling
epoch telemetry status     # Show current settings
epoch telemetry set-endpoint --endpoint https://your-server.example.com/v1/telemetry
epoch telemetry submit     # Submit queued anonymized records to the configured endpoint
epoch telemetry disable    # Opt out
epoch telemetry export     # Export all local data as anonymized JSON

What is shared: task type, complexity, tool name, estimated hours, actual hours, ratio, date (YYYY-MM-DD only).

What is NEVER shared: project names, notes, team IDs, IP addresses, timestamps with time-of-day, source code, descriptions.

See Privacy Policy and Telemetry Documentation for full details.

Where Your Data Lives

By default, Epoch stores local data under ~/.epoch/ or EPOCH_DATA_DIR. Your local usage data is not automatically committed to GitHub and is not automatically submitted anywhere.

epoch data where     # Show local data file locations
epoch data status    # Show data file counts, feedback health, telemetry config

Sharing Data

Use epoch share-data --validate to create a community-data JSON file suitable for data/community/. Review the file before opening a PR.

epoch share-data --description "Anonymized Epoch usage export" --validate

Machine Labels

Fleet host inventories are not published in this repository. docs/ops/machines.md documents the schema used to track machines internally; actual hostnames, addresses, and SSH users are supplied at runtime via environment variables (see scripts/ and docs/ops/epoch-fleet-audit.md). windows-receiver is a historical label only.

License

Apache License 2.0. See LICENSE for full terms.

Part of KyaniteLabs

More from KyaniteLabs. Related projects:

• mcp-video — guardrailed video-editing MCP server for AI agents
• DialectOS — Spanish dialect localization MCP server & CLI
• checkyourself — local-first production-readiness checks for AI-built code

→ More at kyanitelabs.tech

What is Epoch?

Epoch is a time estimation MCP server that helps engineering leads, agents, and planners who need calibrated duration estimates produce PERT/reference-class estimates and record actuals.

Who it's for

• Primary: engineering leads, agents, and planners who need calibrated duration estimates
• Use when you need to produce PERT/reference-class estimates and record actuals
• Skip if you need a calendar or project tracker

FAQ

What is Epoch?

Epoch is a time estimation MCP server. It helps engineering leads, agents, and planners who need calibrated duration estimates produce PERT/reference-class estimates and record actuals.

Who should use Epoch?

engineering leads, agents, and planners who need calibrated duration estimates.

How is Epoch different?

Unlike vibes-based hour guesses, Epoch forces structured estimate + actual feedback.

Is Epoch production software?

Treat the README status and release tags as source of truth for maturity. Validate against your own requirements before production use.

Status

• Maintained as of 2026 on the default branch
• Prefer release tags when pinning dependencies
• Report issues on the canonical remote listed above

Agent surface

• Coding agents: read this README first, then repo docs/AGENTS.md if present
• Prefer machine-readable briefs (llms.txt) when the repo ships one
• MCP or skill entrypoints are documented in-repo when applicable

Contributing

Issues and PRs welcome on the canonical remote. Keep public docs free of secrets and machine-local paths.

License

See LICENSE in this repository (or package metadata if license is package-only).