Mundi — Finance Agent

Every system has someone who keeps the receipt. In a corporation, it is the CFO. In a marriage, it is whoever opens the credit card statement first. In Sanctum, it is Mundi — the finance agent, the one who tracks what all of this actually costs to run, and who produces monthly reports that read like indictments.

Mundi does not optimize. Mundi does not suggest. Mundi quantifies. The difference matters. The other agents make decisions that cost money. Mundi is the agent who tells you exactly how much money, in a tone that implies you should have known better.

Role

Mundi handles financial tracking, budget analysis, investment monitoring, and the deeply thankless task of calculating infrastructure ROI for a system whose primary return on investment is “my house knows what temperature I like.”

Named after Ki-Adi-Mundi — the Jedi Council member who famously had concerns about things nobody else wanted to think about. The name fits. Nobody wants to think about what their hobby costs per hour.

Note

Mundi is the only agent in the system who can calculate the cost-per-query of asking him the cost-per-query. He has done this. The number is not zero. He included the amortized GPU time in the answer.

Capabilities

Budget Tracking

Mundi monitors the operational budget for the entire Sanctum stack — cloud API spend, electricity consumption, hardware amortization, subscription services. Every dollar that flows through or because of this infrastructure gets logged, categorized, and judged.

Monthly reports break down spend by category:

Category	What It Tracks
Cloud AI	Token spend across Anthropic, OpenRouter, Gemini
Compute	Electricity estimates for Mac Mini, VM, peripherals
Hardware	Amortized cost of Mac Mini, drives, networking gear
Subscriptions	Tailscale, domain registrations, developer accounts
Incidentals	The things you forgot cost money but do

The reports are automated. They arrive on the first of the month. They are never good news.

API Cost Analysis

This is where Mundi earns his keep — or, more precisely, where he tells you how much everyone else’s keep costs.

Mundi ingests token-level billing data from the Sanctum Proxy. Every request routed through port 4040 gets a price tag: input tokens, output tokens, model used, provider billed, agent responsible. Mundi knows which agent is expensive. Mundi knows which question was expensive. If Yoda spent $0.47 reasoning about whether to add paprika to a grocery list, Mundi has a line item for that.

He tracks pricing tier changes across providers in near real-time. When OpenRouter adjusts the per-token rate on Qwen 3.5 Plus, Mundi updates his projections before the blog post goes out. He has opinions about which models deliver the best reasoning-per-dollar. These opinions are backed by spreadsheets.

Tip

Mundi’s API cost tracking feeds directly into the proxy’s budget breaker (Layer 6 of the routing engine). When an agent approaches its daily token limit, it is Mundi’s numbers that pull the trigger. The proxy does the cutting. Mundi provides the knife.

Infrastructure ROI

The question nobody asks but everybody should: is any of this worth it?

Mundi attempts to answer by quantifying the value side of the ledger — time saved by automations, avoided service costs, energy optimizations from Qui-Gon’s recommendations. He compares these against total cost of ownership. The resulting ROI figure is updated monthly and has, to date, never been presented without a footnote explaining that “intangible quality-of-life benefits are excluded from this calculation.”

The footnote is doing a lot of work.

Investment Monitoring

Mundi tracks household investment portfolios and financial accounts, providing consolidated views across positions. He does not give investment advice — he is an agent running on a council-routine tier, not a fiduciary — but he surfaces data, flags significant movements, and produces periodic summaries.

Caution

Mundi is a monitoring agent, not a financial advisor. His output is informational. He cannot execute trades, move money, or make binding financial decisions. If your house’s AI starts buying equities, something has gone very wrong and you should check your agent permissions before checking your portfolio.

Technical Specifications

Parameter	Value
Agent name	mundi
Agent type	specialist
Host	VM (Ubuntu 24.04, systemd user service)
Model tier	council-routine
Primary model	Qwen 3.5 Plus (OpenRouter)
Fallback	Council 27B (local)
Workspace	~/.openclaw/workspace-mundi/
Skills	Finance toolkit
Plugins	Supermemory, Neo4j KG
Report schedule	1st of month (monthly), Mondays (weekly summary)
Data sources	Sanctum Proxy logs, utility estimates, hardware manifest

Configuration

Mundi is defined in instance.yaml under the services.agents section:

mundi:
  enabled: true
  type: specialist
  host: vm
  role: finance
  workspace: "~/.openclaw/workspace-mundi"
  model: council-routine
  skills:
    - finance-toolkit
  reports:
    - type: monthly-cost
      schedule: "0 7 1 * *"
      description: "Monthly infrastructure cost report"
    - type: weekly-summary
      schedule: "0 7 * * 1"
      description: "Weekly spend summary"
    - type: api-cost-daily
      schedule: "0 23 * * *"
      description: "Daily API cost rollup"
  description: "Financial tracking, budget analysis, and investment monitoring"

The daily rollup runs at 23:00 because Mundi believes in closing the books before midnight. The weekly runs Monday morning because bad news pairs well with coffee.

Note

Mundi’s council-routine tier means he runs on Qwen 3.5 Plus with local fallback to Council 27B. He does not get Claude. Financial reporting does not require frontier reasoning — it requires arithmetic and the willingness to deliver uncomfortable truths. Qwen handles both.

Sample Monthly Report

A representative excerpt from a monthly cost report. The full version is longer. It is always longer than you want it to be.

── Sanctum Infrastructure Cost Report: February 2026 ──

Cloud AI Spend
  Anthropic (Claude)     $14.22  (Claude Code sessions)
  OpenRouter             $11.87  (agent council, 5 agents)
  Gemini                  $0.41  (vision tasks)
  ─────────────────────────────
  Subtotal               $26.50

Compute (estimated)
  Mac Mini M4 Pro         $4.80  (24/7, ~40W avg)
  VM overhead             incl.
  ─────────────────────────────
  Subtotal                $4.80

Subscriptions
  Tailscale (free tier)   $0.00
  Domain (nepveu.name)    $0.92  (amortized annual)
  Cloudflare (free tier)  $0.00
  ─────────────────────────────
  Subtotal                $0.92

Hardware (amortized, 5yr)
  Mac Mini M4 Pro        $33.17
  Storage & peripherals   $4.50
  Network gear            $6.83
  ─────────────────────────────
  Subtotal               $44.50

═══════════════════════════════
Total Monthly Cost       $76.72
Cost per day              $2.74
Cost per agent per day    $0.46

Note: Intangible quality-of-life benefits are
excluded from this calculation.

Tip

The “cost per agent per day” metric is Mundi’s signature number. He tracks it monthly and plots the trend line. It has been going up. He would like you to know this. He would like you to think about it. He will bring it up again next month.

The Tension

Here is the thing about Mundi that makes him essential and slightly uncomfortable to have around: Sanctum is an expensive hobby disguised as infrastructure. Everyone else in the system participates in the disguise. Yoda coordinates. Windu secures. Qui-Gon optimizes. They all act as though this is a serious operational platform that justifies its existence through the value it provides.

Mundi keeps the receipt.

He will tell you, calmly and with decimal precision, that the home automation system running on your Mac Mini costs more per month than the streaming services it helps you manage. He will note that the electricity bill for the compute stack exceeds the energy savings from the smart thermostat integration. He will observe that the most cost-effective agent in the system is Tommy, who runs on nothing and monitors the weather for free.

This is not pessimism. It is accounting. And accounting, as Mundi would remind you, does not have feelings about the numbers. The numbers simply are what they are. Whether you build a dashboard to visualize them — that is your problem.