Skip to content
Foundry

foundry-economics

foundry-economics puts numbers on decisions: what work will cost, what a system costs to run, what running the AI itself costs, and whether the thing is worth building at all.

Every estimate it produces is a three-point range with the assumptions written down, because a single number presented without assumptions is a guess wearing a suit.

Install

Section titled “Install”
Terminal window
/plugin install foundry-economics@foundry

Requires foundry-core, which is installed automatically as a dependency.

When to install it

Section titled “When to install it”
  • Someone asks “how much will this cost”, “how long will this take”, “what is the TCO” or “are we on budget”.
  • The cloud bill is rising and nobody can attribute it to a feature or a request.
  • You want to know what Claude Code itself is costing, which model an agent should use, or whether prompt caching is worth it.
  • A grant budget, a public-call structure, timesheets or milestone reporting has to be prepared.

When not to use it

Section titled “When not to use it”
  • These are engineering estimates, not accounting. Nothing here is an audited financial statement or tax advice.
  • It does not sequence work — that is roadmap-planner in foundry-pmo.
  • Cloud cost analysis needs billing data. Without it, finops-analyst reasons about architecture, not about your actual bill.

Agents

Section titled “Agents”
Agent What it does Model Effort
cost-engineer Project cost engineering: decomposes work, produces three-point estimates and answers “how much”, “how long”, “what is the TCO”, “build or buy”. opus high
business-case-analyst Business case and financial plan: whether to build something, break-even, the numbers a funding or board paper needs. opus high
finops-analyst Cloud and infrastructure run-cost analysis: why spend is up, what a request costs, whether reserved capacity is worth buying. opus high
funding-analyst Grants and public funding mechanics: grant budgets, structuring a project for a public call, timesheets and evidence, milestone reporting. opus high
ai-cost-controller The cost of running AI itself: what Claude Code is costing, which model an agent should use, prompt-cache economics, token budgets per workflow. sonnet medium

Skills

Section titled “Skills”
Skill When it fires
estimate-project A structured estimation session: decompose the work, collect three-point ranges, compute PERT expected value and p50/p80/p95, sanity-check.
tco-model A multi-year total cost of ownership model over a stated horizon, with discounting and the cost lines teams routinely forget.
business-plan A financial plan: P&L projection, cash flow, break-even, unit economics, NPV/IRR at a stated discount rate, with scenarios and a sensitivity table.
ai-spend-report A report of AI and token spend from real metrics — per-agent and per-feature attribution, retry waste, prompt-cache economics, token budgets.
budget-tracking Setting up and running budget-versus-actual tracking with earned-value variance analysis, forecast at completion, and an agreed escalation threshold. setup once, review --period YYYY-MM thereafter.

tco-model and business-plan take an explicit --horizon and --rate, so the discount rate is a stated input rather than a hidden assumption.

Output contracts

Section titled “Output contracts”
Agent Input Output
cost-engineer plan.v1 estimate.v1
business-case-analyst estimate.v1 estimate.v1
finops-analyst estimate.v1 estimate.v1
funding-analyst estimate.v1 compliance-check.v1 — funding-programme obligations assessed as controls
ai-cost-controller estimate.v1, optionally, plus the project’s own instrumentation estimate.v1, one item per attribution unit (feature, agent or workflow)

estimate.v1 requires scope, at least one items entry with optimistic, likely and pessimistic, and at least one assumptions entry. An estimate with no stated assumptions cannot be written to the blackboard — the schema rejects it.

The AI-cost loop

Section titled “The AI-cost loop”

ai-cost-controller and ai-spend-report read what foundry-core records. Every memory_search, memory_write, blackboard_write, gate block, subagent return and session end is appended to .foundry/metrics/events.jsonl, and the token_report MCP tool and the foundry tokens command summarise it. That makes “which agent is expensive” a measurement rather than an opinion.

Limits

Section titled “Limits”
  • Token figures are estimated at roughly four characters per token. They are consistent enough to compare agents against each other; for billed amounts use /cost and /usage.
  • .foundry/metrics/ is gitignored by foundry init, so the history is local to a machine unless you deliberately collect it.
  • Currency defaults to EUR in estimate.v1 and risk.v1 (impactEur). Other currencies are expressible in estimate.v1 via currency, but the risk schema’s exposure field is EUR-named.
  • Funding rules differ per programme and per call. funding-analyst produces evidence structure and budget mechanics; it does not certify eligibility.