io.github.labham-llc/agentcents MCP Server

A local LLM cost proxy that forecasts each call and recommends the cheapest model that clears the task’s capability bar. It provides cost estimates for call or agent workflows, including honest ranges rather than false point estimates, and compares options visually.

🛠️ Key Features

Forecasts call costs and estimates agent/workflow costs
Recommends the cheapest model that meets the capability bar
Shows model comparisons as a Mermaid diagram
Uses honest ranges for pricing estimates
Offers consult mode with a bundled snapshot of model prices and Intelligence Index scores

🚀 Use Cases

Selecting a cost-optimized model for a specific LLM call or agent workflow
Comparing candidate models for predicted cost and capability fit

⚡ Developer Benefits

Predictable cost guidance across candidate models
Clear model comparison output via Mermaid diagram
“Consult mode” option without installing anything

⚠️ Limitations

Live mode details are only partially present in the provided data (“Live mode:” appears, but the description is truncated).

agentcents — LLM cost consulting skill for Claude

A Claude Skill that turns Claude into a neutral LLM cost advisor: estimate what any call or agent workflow will cost across candidate models, get the cheapest model that clears the task's capability bar, and see the comparison as a Mermaid diagram — with honest ranges, never false point estimates.

Works two ways:

Consult mode (no install): answers from a bundled, hand-verified snapshot of current model prices and Intelligence Index scores.

Live mode: if the agentcents proxy is running locally, the skill uses your real ledger — calibrated forecasts, spend, and accuracy instead of priors.

Install

Claude.ai / Desktop: Settings → Capabilities → Skills → Add → upload this folder (or the packaged .skill).

Claude Code: /plugin marketplace add labham-llc/agentcents-skill then /plugin install agentcents@labham-skills

Then just ask things like:

what would it cost to classify ~500 support emails a day, about 200 tokens each?

I'm building a 5-step research agent — which model should each step use?

What's in the box

code

SKILL.md                 — the consulting method (measure, range, bar, pick, flag instability)
assets/pricing.json      — price + capability snapshot (see _meta.verified_utc)
references/local-api.md  — live-mode reference for a locally running agentcents

Snapshot verified: 2026-07-15 (see _meta.verified_utc) — refreshed with each agentcents release.

The product behind it

agentcents is a local proxy that meters, forecasts, and budget-enforces LLM API calls — and grades its own forecasting accuracy on a calibration page. pip install agentcents. The skill is free; so is the agentcents free tier.

io.github.labham-llc/agentcents