Work: mcp-unifi, Writ Protocol, Open Model Arena, DewFront and More

Flagship products

Built in the open, run in production.

Each one started as a problem we had ourselves. Each one ships with tests, CI, a security policy and documentation, because that is the standard we sell.

01 MCP server MIT Open source

mcp-unifi

Run a UniFi network in plain English, with guardrails an assistant cannot talk its way past.

A safety-first MCP server that lets Claude, or any MCP client, manage a self-hosted UniFi network: VLANs, firewall rules, WiFi and switch ports. Opt-in modules add Protect for cameras and read-only Access for doors. The rules live in the server, not the prompt. Every destructive call can be previewed with a dry run, multi-step changes roll back on partial failure, and every call lands in an audit log with secrets scrubbed.

100+ tools across UniFi Network, Protect and Access
Dry-run previews, composite rollback, JSONL audit log
Read-only mode, multi-site, cosign-signed images with an SBOM
Docker, Helm, Claude Desktop bundle, and the official MCP Registry

GitHub Docs Why we built it

mcp-unifi example session

> put the smart speakers on their own network, away from the laptops

dry_runcreate_iot_network: 3 changes predicted

previewvlan 30 iot, wlan iot, 1 drop rule to the main LAN

applyapproved and applied by the owner, 3 of 3

auditlogged to audit.jsonl, secrets scrubbed

02 Agent protocol Apache 2.0 Draft spec v0.1

Writ Protocol

Hand an AI agent a limited slice of your authority, and get back a signed record of what it did with it.

Writ lets one agent delegate a narrower piece of its authority to another, which can narrow it again, and brings back a signed receipt from every hop. Anyone can check the chain offline without asking a server. It answers the questions agent protocols leave open: can the limit be trusted, what actually happened, and can it be undone.

Draft specification plus Go and Python implementations that check each other
288 conformance vectors and 25 multi-step scenarios
writ-hook puts a Claude Code session under a signed grant
Apache 2.0, built for an outside implementer to break

GitHub Read the spec

writ example session

> book the trip, spend at most $600

grantA to B: book travel, max $600

narrowB to C: charge card, max $589

tallyC signed: charged $589, ok

verifyB and C receipts valid, chain intact

03 Evaluation app MIT Open source

Open Model Arena

Find out how the model on your own hardware compares to the API you pay for, on your own prompts.

Two models get the same prompt. You read both answers without knowing which wrote which, then vote. An ELO leaderboard tracks the results, and per-response cost tracking shows what each answer cost. It connects to any OpenAI-compatible endpoint, local or cloud, and runs as one container with one YAML file.

Blind battles with live streaming and an ELO leaderboard
Per-response cost tracking, categories and a vote audit log
Audience vote: a room full of phones picks the winner
Ollama, vLLM, LM Studio, gateways and cloud APIs side by side

GitHub Docs

open-model-arena example session

> summarize this support ticket for the on-call engineer

model alocal 8B on a Mac mini, $0.000

model bcloud API, $0.004

voteA wins, blind

elolocal 8B +16, now 1,032

04 Game engine MIT Live on two sites

Open Setlist Stash

A prediction game that reads any band’s setlist data over MCP. Two live communities, one image, zero forks.

The game engine holds no setlist data of its own. Every read goes out over the Model Context Protocol to a server you point it at, so the same container runs Tweezer Picks for Phish fans and Wappy Picks for Umphrey’s McGee fans. Only the environment config and branding differ.

Runs tweezerpicks.com and wappypicks.com from one published image
Data arrives over MCP, so a new band needs a data source, not a fork
FastAPI and Postgres, mypy-strict, Trivy-scanned
Self-hostable for any music community

GitHub tweezerpicks.com wappypicks.com

open-setlist-stash example session

> score tonight’s picks against the setlist

mcpget_show(date) from mcp-phish

resolveset 1 and set 2 locked

scorestandings updated in one tick

tenantsame image serves wappypicks.com

Showcase

DewFront. A weather app that makes the decision for you.

Most weather apps hand you a table and leave the thinking to you. DewFront leads with dew point, the reading that actually tells you how the air will feel. It weighs dew point against temperature, rain chance and wind to give one verdict on whether to open the windows tonight, and names the best two hour window today for getting outside.

The overnight rule, live

Should the windows be open tonight?

8 PM open

Dew point 60°F or below
Cooler outside than the 70°F target
At or above the 50°F cold floor
Rain chance under 35%
Wind under 20 mph

Verdict

Open until 6 AM

At 6 AM it drops to 49°F, under the 50°F cold floor.

Four made-up nights, run through the rule as the case study documents it: an hour is open only when all five conditions hold, and the verdict comes from the longest unbroken run. The app itself goes further, from shoulder seasons to dry winter air.

DewFront on a phone: the Now screen leads with the temperature, the dew point and its comfort word, then a plain sentence about how the air will feel — The Now screen on a phone

01
Every reading says where it came from.

Each field takes the best fresh source: your own weather station, then the nearest National Weather Service station, then a forecast model. Measured readings are marked, and a reading that has aged out is shown struck through rather than hidden.
02
It shows its own doubt.

Three independent global models sit side by side for the next three days, and the app says whether they agree. One confident number would hide something you would want to know.
03
One rule, used everywhere.

The server and the browser run the same decision functions, compiled into both, with the clock passed in. There is no copied threshold to drift, and any night can be replayed in a test.
04
Tests that prove what is missing.

The tests worth pointing at assert absence: relative humidity never appears on the windows card, and switching to Celsius leaves no Fahrenheit anywhere on the screen.

2,501
unit tests: 97.55%
statement coverage: 91.31%
branch coverage: 55
end-to-end tests

Test figures from the public case study, measured 2026-09-19. The production source is private; the case study covers what it does, how it is built, and the decisions behind it, including the ones that were wrong first.

Visit dewfront.com ↗ Case study

Showcase

All Terrain Society. A riding crew’s record, mapped by an agent.

All Terrain Society puts a riding crew’s trips and spots on a live map, with each ride’s route drawn underneath and a GPX file to download. An agent pipeline publishes it from strava-mcp-vault, our open source MCP server for Strava, and nothing ships until the site’s checks pass.

From Strava to the map every refresh, in this order

01 Strava Where rides are recorded
- 100 requests per 15 min
- 1,000 a day
02 strava-mcp-vault Our open source MCP server
- SQLite cache, API only on a miss
- tokens renewed before they expire
03 Sync agent Pulls the new rides
- rebuilds spots, photos, GPX
- new places wait for a name
04 Gates Checks before anything ships
- privacy test
- a real browser sees the map draw
05 Review Stops for a person
- nothing deploys unseen
06 allterrainsociety.com Live
- deploys, then rechecks the live map

One ride, as the map shows it GPX ↓

Fall Mountain Bike Ride Shindagin Hollow · Sep 30, 2026 · 7.1 mi · 699 ft climbing

All Terrain Society's Where We Ride page on desktop: totals for spots, rides, miles and climbing above a map of New York and Vermont, with green pins numbered by the rides at each spot — Where We Ride · every spot, pinned by hand

Three promises the server keeps

Each one is pinned by a test that fails if the promise is removed.

01
Only public rides.

A ride is published only when Strava marks it visible to everyone. A missing or unknown setting is withheld, so a change on Strava’s side fails closed.
02
Nothing named by a machine.

A spot appears only after someone names it. Unnamed rides are counted, never placed: auto-naming would have put a neighborhood with 59 ride starts on the map first and largest.
03
No start points, ever.

The export never emits a recorded start coordinate. Only route lines go out, and a spot’s pin is the coordinate a person chose.

187
rides on the live map: 27
riding spots: 10
GitHub stars, all outside: 5
forks, all outside

Ride counts read from the live site on 2026-10-04, stars and forks from GitHub, every one of them from someone else's account. strava-mcp-vault is MIT licensed, and an outside contributor has had two pull requests merged. It is the same shape as our Read-only MCP connector package.

Visit allterrainsociety.com ↗ strava-mcp-vault

Open source

The tools behind the tools.

MCP servers, a shared library, playbooks, tools and incident write-ups. If we use it to run client work, there is a good chance you can read it.

Sites we build and run

Our own properties.

Not client work. These are ours: built, hosted, patched and paid for by us.

DewFront Weather, decided

All Terrain Society Ride map, fed by MCP

Tweezer Picks Phish setlist game

Wappy Picks Umphrey’s setlist game

Bands That Jam Where it started, 2007

Pulse of Ithaca Local podcast

Web work · since 2009

Where it started.

Sites we have built, hosted or maintained for festivals, artists and businesses across New York and beyond. We still build, host and look after websites, on their own or as part of a bigger project, to the same standards as everything above. See the tiers and prices.

Clients we have worked with

See services

Have a system you want an agent to check?

No sales pitch and no discovery call that is really a sales call. A straight answer about what we can build, how we would keep it safe, and what it costs, quoted in writing.

Get in touch pete@brooksnewmedia.com

Products we run. Work we have shipped.

Built in the open, run in production.

mcp-unifi

Writ Protocol

Open Model Arena

Open Setlist Stash

DewFront. A weather app that makes the decision for you.

Every reading says where it came from.

It shows its own doubt.

One rule, used everywhere.

Tests that prove what is missing.

All Terrain Society. A riding crew’s record, mapped by an agent.

Only public rides.

Nothing named by a machine.

No start points, ever.

The tools behind the tools.

Our own properties.

Where it started.

Have a system you want an agent to check?