AI can write code.
But it can't tell time.

Epoch gives AI agents the ability to estimate timelines, calculate costs, and predict deadlines — with one command.

Terminal
$ claude mcp add epoch -- npx @puenteworks/epoch

The problem no one is solving

AI agents are incredibly powerful. They write code, debug systems, and automate workflows. But ask them "how long will this take?" and they guess — badly.

The Problem

Your AI agent confidently says "this will take 2 hours." Two days later, you're still waiting.

Large language models have no concept of time. Positional encodings represent token sequence position, not real-world duration. Self-attention provably cannot model counting behaviors.

  • AI agents estimate in token budgets, not minutes
  • Events 10 tokens apart have identical encodings whether 1 second or 1 hour apart
  • LLMs replicate and amplify human optimism bias through agentic overconfidence
The Solution

Epoch gives AI grounded estimates using 40+ years of software engineering research, packaged into 24 tools any AI can call.

Instead of hallucinating timelines, your agent calls a function. The function returns a data-driven estimate built on COCOMO II, PERT analysis, reference class forecasting, and Monte Carlo simulation.

  • Structured external tools — not prompt hacks
  • 24 MCP tools covering time, dates, estimation, analytics, and cost
  • One command to install. Works with Claude, Cursor, VS Code, and any MCP client

Research-backed: A UPenn negotiation study showed 4% deal closure under wall-clock time limits vs 99% under turn-based limits. Injecting explicit time state into prompts produced an 8x improvement in negotiation outcomes. Epoch makes time legible to LLMs through structured external tools.

See it in action

A real interaction. No cherry-picking, no edits.

Claude + Epoch

How long will a 15K LOC project take with a team of 5?

Claude calls cocomo_estimate, pert_estimate, critical_path

Based on COCOMO II estimation:

Effort: 45 person-months Duration: 8.3 months Risk Level: medium Estimated Cost: ~$340K

Confidence: 72% | Method: COCOMO II + PERT blend

5 layers of temporal intelligence

From basic time operations to Monte Carlo simulation. Each layer builds on the last, giving your AI progressively deeper temporal reasoning.

L1

Time & Dates

Current time in any timezone, timezone conversion, duration parsing, date arithmetic

get_current_time convert_timezone parse_duration time_math
L2

Business Days & Holidays

Working day calculations, holiday awareness (US, UK, FR, DE, JP), urgency classification

add_business_days count_business_days
L3

How Long Will It Take?

PERT three-point estimation, COCOMO II (LLM-adapted), sprint velocity, critical path method

pert_estimate cocomo_estimate sprint_forecast critical_path
L4

Getting Better Over Time

Reference class forecasting, estimation accuracy tracking, historical data calibration

reference_class_estimate calibrate_estimates
L5

Cost & Risk

Monte Carlo simulation, token-to-time bridge, planning fallacy correction, team velocity calibration

token_time_bridge monte_carlo_schedule

What people are saying

Early adopters are using Epoch to give their AI agents temporal superpowers.

AK
AI Engineer
Y Combinator Startup

"We went from AI estimates that were off by 10x to estimates backed by real data. Epoch changed how our entire team plans sprints."

PL
Platform Lead
Enterprise SaaS

"The Monte Carlo simulation alone is worth it. Our PM team finally trusts the AI's deadline predictions."

SR
Senior Researcher
AI Safety Lab

"This is the right approach. External tools with grounded data, not prompt engineering. Epoch is how AI time estimation should work."

Testimonials are illustrative. Have a story? Share it on GitHub.

All 24 tools, one API

Zod-validated schemas, .describe() on every field, readOnlyHint annotations. Accessible through MCP, CLI, and REST.

Tool Description
get_current_timeCurrent time in any IANA timezone
convert_timezoneConvert ISO-8601 between timezones
parse_durationParse "2h30m" into structured seconds
time_math6 operations: add_days, diff, format_duration, add_hours, add_minutes, add_weeks
add_business_daysAdd N working days, skipping weekends and holidays (5 countries)
count_business_daysCount working days between two dates
pert_estimateThree-point PERT Beta distribution: E = (O+4M+P)/6
cocomo_estimateCOCOMO II with 5 LLM-adapted cost drivers
sprint_forecastSprint velocity forecasting with variance-based pessimistic estimate
critical_pathCPM with forward/backward pass and merge bias
monte_carlo_scheduleMonte Carlo simulation with seeded PRNG, P10-P95 percentiles
reference_class_estimateReference class forecasting with industry correction factors
calibrate_estimatesTeam-specific correction factor calibration from historical data
token_time_bridgeMap token budgets to wall-clock time for 12 LLM models
token_cost_estimateToken cost estimation with input/output breakdown
compare_modelsSide-by-side comparison across 12 LLM models
accuracy_trendSliding-window MAPE tracking over time
schedule_riskConfidence intervals using team MdAPE
record_actualSubmit actual hours for a previous estimate
feedback_healthComprehensive feedback data quality report
896
Tests Passing
97%
Statement Coverage
24
MCP Tools
12
Model Calibrations

Gets better the more you use it

Every estimate generates a feedback token. Record actual hours and Epoch auto-calibrates correction factors across task types, complexity levels, and tools.

369
Matched Pairs
6
Projects
8
Task Types
Auto
Correction

Transparent by Design

Every data point is documented. Every record is anonymized. Telemetry is off by default.

Off by Default

Telemetry requires explicit opt-in. Preview exactly what will be shared before enabling.

Zero PII

No names, emails, project names, or company data. Dates truncated to day-only.

Self-Hostable

Configurable endpoint. Run your own telemetry server. Full API contract documented.

Three surfaces, one API

All 24 tools are accessible through MCP, CLI, and REST API. Use whichever fits your workflow.

MCP Server

Stdio transport for Claude Code, Cursor, VS Code

claude mcp add epoch -- npx @puenteworks/epoch

CLI

Scriptable subcommands with JSON or table output

epoch pert-estimate -o 2 -m 4 -p 12

REST API

HTTP server with OpenAPI spec and AI discoverability

epoch serve --port 3000
/.well-known/ai-plugin.json
AI manifest
/llms.txt
LLM docs
/openapi.json
OpenAPI 3.1
/health
Status

Up and running in 30 seconds

No config files. No API keys. One command.

1

Install Epoch

$ claude mcp add epoch -- npx @puenteworks/epoch
2

Ask any time question

"How long will it take to build a 15K LOC project?"
"What's the PERT estimate for a 3-month sprint?"
"How many business days until March 15?"
3

Get grounded, data-driven estimates

{
  effort: "45 person-months",
  duration: "8.3 months",
  risk: "medium",
  confidence: "72%"
}

That's it. Three steps.

View on GitHub

Built on solid foundations

TypeScript 5.8
Strict + noUncheckedIndexedAccess
Zod 3.24
Schema validation with .describe()
MCP SDK 1.12
Model Context Protocol
Vitest 3.x
896 tests, v8 coverage