fairhaul-agents
Processes DHL parcel invoice data to identify late deliveries, duplicate charges, surcharges, and address corrections, and prepares refund claims for approval.
Processes FedEx parcel invoice data to identify late deliveries, duplicate charges, surcharges, and address corrections, and prepares refund claims for approval.
Processes UPS parcel invoice data to identify late deliveries, duplicate charges, surcharges, and address corrections, and prepares refund claims for approval.
Processes USPS parcel invoice data to identify late deliveries, duplicate charges, surcharges, and address corrections, and prepares refund claims for approval.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Fairhaul freight agents
Thirty small MCP agents for freight and shipping cost. Each one owns one job, such as auditing a freight bill, building the month-end accrual, settling carrier pay or checking a customs entry, and cannot act outside it. They read the files you already export from your TMS, ERP, WMS, carrier portals and EDI provider, write reports on your machine, and wait for a named person before anything that moves money, sends a message or changes a live system.
MIT licence · Node 20 or newer · local files only · no credentials · no dependencies
Also in: Deutsch · Français · Español · Português (Brasil) · Italiano · 日本語 · 简体中文 · 한국어 (translated by Claude; native review pending)
Try one
npx -y fairhaul-freight-bill-audit demo # freight bill audit on fictional data
npx -y fairhaul-freight-accruals demo # month-end freight accrual on fictional dataEvery agent has a demo that runs on fictional files, an eval that checks it against its own test cases, and an MCP server (npx -y <agent> mcp --root <folder>) for Claude Desktop, Claude Code, Cursor and other MCP hosts. Each agent's README lists its tools.
Related MCP server: noxgild-cli
The agents
Audit what carriers bill
Agent | What it does | Waits for a person before |
Truckload and LTL bills (CSV, Excel or X12 210) against your contract rates, fuel table and accessorial terms; duplicates and math errors | Sending a dispute or approving payment | |
UPS, FedEx, DHL and USPS invoice data: late deliveries, duplicates, surcharges, address corrections | Submitting refund claims | |
Detention, layover and lumper charges against your dock times and free time | Sending a dispute | |
Every fuel surcharge against the carrier's fuel table and the diesel price for the right week | Sending a dispute | |
Where accessorial charges come from (liftgate, residential, limited access) and the address or order fixes that stop them | Changing addresses or order rules | |
LTL weight and class corrections against your declared dimensions and density | Sending a challenge | |
Years of freight payments for invoices paid twice, under variant numbers or two vendor records | Sending a recovery request | |
Ocean invoices line by line against your service contract, surcharges and validity windows | Sending a dispute | |
Air freight invoices: chargeable weight, rate breaks and surcharges | Sending a dispute | |
Every delivered or invoiced shipment matched to its proof of delivery; what to chase and what to hold | Sending POD requests | |
Container free time at the port and with the line, demurrage and detention exposure, charges to dispute | Sending a dispute | |
Customs entry lines against commercial invoices: value, quantity, origin, duty arithmetic, codes used inconsistently. No classification advice | Sending a note to the broker |
Plan, buy and run freight
Agent | What it does | Waits for a person before |
RFP bids scored lane by lane with share caps, incumbents, backups and capacity | Sending awards | |
Your shipment history re-rated under every contract you hold, lanes ranked by money left on the table | Changing the routing guide | |
Spot quotes on one all-in basis (fuel, accessorials, expiry, transit) | Tendering the load | |
Spend by lane and mode, cost per hundredweight trends, consolidation opportunities | Sending a consolidation plan | |
Carrier scorecards on on-time, tender acceptance, claims and cost | Sending a scorecard | |
Warning signs of fraud and double brokering in the carrier master and booked loads | Sending a verification request | |
Open loads with stale or late tracking, at risk of missing delivery | Sending delay notices | |
A day of dock appointments across doors, crews and hours; no-shows and waits in past appointments | Publishing the schedule | |
Every parcel's carton against its contents, dimensional weight saved by a better fit | Changing packing rules | |
X12 204, 210, 214, 990 and 997 files read, converted and reconciled | Never transmits EDI |
Close the books and cost the business
Agent | What it does | Waits for a person before |
Month-end freight accrual with the basis of every estimate and a draft journal | Approving the journal for posting | |
Freight bills split across orders, customers, SKUs and cost centers, with a draft reclass journal | Approving the journal for posting | |
Standard freight in item costs against actual freight per unit; stale standards and proposals | Loading new standards | |
Freight charged to customers against what it cost; collect orders you paid, free-shipping leaks | Changing customer terms | |
For brokers: carrier pay against rate confirmations, quick pay and notices of assignment | Releasing a payment batch | |
Loss, shortage and damage claims with liability caps and filing deadlines | Filing a claim | |
Returns from authorisation to credit, and return freight billed twice | Sending a carrier dispute | |
Shipment emissions estimated from weight, distance and the factors you choose. Estimates only | Sending a customer statement |
How they stay safe
Allowlist per agent. Each agent can call only its own tools. Anything else is blocked, including tools a host model invents.
Masked before the model sees it. Names of parcel and return recipients, people who signed for a delivery, contacts and drivers become pseudonyms (
CU-3f2a9c1d,ST-8cea1a2d); phones, e-mails and street addresses become placeholders, in every result an AI model reads. Carrier and customer company names, load and invoice numbers, dates and amounts are sent as they are (--pii strictalso hides company names). The files written on your machine keep the real values.Refused tools. Each agent names the live actions near its job (paying, emailing, changing the TMS, posting to the ledger, filing with customs) and blocks them, so a model asking for them gets a clear refusal.
One named click. Approving a payment, a send, a production change or a journal waits for a named person, even in Auto mode. The approval is tied to the sha256 of the reviewed draft; if the draft changes, the approval no longer matches.
Audit log. Every call is written to a hash-chained log under
.fairhaul/, signed with an Ed25519 key when you configure one.Kill switch.
<agent> killstops every tool until<agent> unkill.Hostile input tested. Spreadsheet formulas, HTML, path tricks and text aimed at the model are planted in every agent's data in CI (
npm run hostile).
Use with Claude Code
This repository is a Claude Code plugin marketplace. Each agent is a plugin that starts its MCP server with npx:
/plugin marketplace add <path or git URL of this repository>
/plugin install fairhaul-freight-bill-audit@fairhaul-freightThe plugins start npx -y <agent>@<version>, so they need that version on npm. Until then, add an agent from a local checkout:
claude mcp add fairhaul-freight-bill-audit -- node /path/to/fairhaul-agents/packages/fairhaul-freight-bill-audit/bin/fairhaul-freight-bill-audit.js mcp --root .Develop
npm ci
npm test # every package's tests and evals
npm run evals # every agent's eval cases, with totals
npm run accuracy # planted problems found, findings on clean data, mutations (docs/accuracy.md)
npm run model-view # what each agent sends to the AI model (docs/model-view.md)
npm run settings # every day count, rate and limit an agent applies (docs/settings.md)
npm run realdata # checks on public real data (download first with realdata:download)
npm run intl # every spreadsheet eval case in 8 languages
npm run hostile # planted hostile inputs in every agent
npm run mcp:check # the official MCP SDK client against every agent
npm run readme:check # generated README sections are up to date
npm run bench # speed at 1,000 and 10,000 rows
npm run plugin:check # plugin files match the packages
npm run release:check:offline # pack contents, licences, docs and versions
node scripts/new-agent.mjs fairhaul-<what-it-does> "<description>" kw1,kw2 <module>CI runs on Node 20, 22 and 24, on Linux and Windows.
How they are tested
Each agent has hand-written tests for its own rules, the same contract tests as every other agent (arguments, files outside --root, Review mode, the kill switch, draft approval, repeatable results) and 14 to 22 eval cases on fictional files. Every spreadsheet case is run again as a German, French, Spanish, Brazilian, Italian, Japanese, Chinese and Korean export, and hostile data is planted in every agent's files. In total: 953 tests and 551 eval cases (final pass, 19 Nov 2026, Windows and Node 24; the CI covers Node 20 and 22). REVIEW-ver3.md lists what changed in ver 2 and the bugs the tests found.
Limits
Every agent works from the files you give it. None of them logs in to a carrier, TMS, ERP, bank or government system.
Findings and estimates are only as good as those files. Each report says what it assumed.
Not legal, tax, customs or accounting advice. People decide and act.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
License checks, signed rights proof, keyless pre-call checks for MCP servers, and the agent store
MCP gateway with runtime security policy, tool-call-level control, and audit of agent actions.
Run and manage H Company's Computer-Use Agents from any MCP client.
Use a person's computer after they approve it there: files, Terminal, browser, local MCP.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables an MCP client to safely access selected local files and trusted executables with policy control, audit, and rollback via a Windows control center.Apache 2.0

noxgild-cliofficial
AlicenseNot gradedqualityCmaintenanceConnects MCP-compatible AI clients to an explicitly authorized Windows, macOS, or Linux computer for controlled Machine, Browser, and Desktop execution, including files, shell commands, native apps, and local browser sessions. Also offers a separate hosted Cloud Browser for browser tasks that should run outside the user's own workstation.7 npmApache 2.0- AlicenseNot gradedqualityBmaintenanceEnables AI agents to access the local machine over MCP for terminal commands, file operations, code search, process management, and reading or editing PDF, Excel, and Word documents, with no telemetry.394 npm1MIT
- AlicenseAqualityBmaintenanceEnables local, offline inspection of computer activity by collecting visible process, connection, and automatic-launch metadata and letting MCP clients explain rule matches and process context. It is read-only toward processes and system settings and offers review suggestions without making changes.8MIT