Software that ships itself

Last month, 500 of 504 releases shipped with no human on them.

500 of 504 production releases in September were made by our AI agents, not a person. That's what two and a half years and 105 billion tokens of real client work bought us — and it's the platform we now hand to customers.

Book a demo →See the platform

Built with fabric

Context Architecture

105 billion tokens processed at 95% cache efficiency.

Context architecture engineered for deterministic efficiency at enterprise scale.

How We Achieved This →Latest Dev Report →

CONTEXT

CACHE

CONTEXT INPUT

4.9%

Written once, cached

CACHE REUSE

95%

Massive scale efficiency

105B

TOKENS PROCESSED

95%

CACHE HIT RATE

312.7M

OUTPUT TOKENS

4.9%

CONTEXT WRITTEN

The Architecture

·Large 1M-token contexts written once, then read hundreds of times

·95% of all tokens served from cache, not recomputed

·Standards inherited by every agent instead of pasted per prompt

·The stable part of a prompt kept byte-identical, so it stays cacheable

What It Unlocks

·Retries cost almost nothing, so an agent can fix itself

·16 agents on one codebase without the bill scaling with them

·500 of 504 releases shipped with no engineer in the loop

·Proven on 105B tokens of real client work, not a benchmark

The Result

Software that ships itself.

We engineered ourselves out of the loop.

MACE INNOVATIONS

* Token and cache figures are actual organization-wide Anthropic API consumption over the trailing twelve months; cache efficiency = cache-read input tokens ÷ total input tokens. Development metrics (commits, active repos, lines) are pulled from the GitHub API. Every number on this site comes from source data — nothing is fabricated.

SEPTEMBER 2026 DEVELOPMENT

4,146 Commits. 500 of 504 Shipped by Agents.

Our biggest month yet — and 500 of 504 production releases still went out with no human on them. Four had a person on them, up from one in August, and we say so.

Real data — not marketingPulled live · GitHub + Anthropic APIs

⚡

Total Commits

📦

Active Projects

🚢

Publishes to Prod

📈

Lines Added

Work Breakdown

2,758 commits

General Updates1,15342%

Iterative Refinement71126%

New Features42315%

DevOps / Config2599%

Documentation2098%

Code Style30%

The agents close the loop. With AI the first pass is a starting point, so most “fixes” are refinements rather than bugs — and in September the refining kept not going through us. 500 of 504 production releases were shipped by an agent; four were a human merging by hand (August: one).

Stack & Focus

Platform

FabricJerryNeoMulti-Agent Platform

Delivery

One-Button Go-LiveVercelPreview Envs

Lending

Encompass / ICEMeridianLinkOptimal BlueLoanscapeFigure HELOC

Data

Experian CreditResWare Title & Fees

Comms

Feather Voice AITwilio SMSEmailSecure Delivery

UI Kit

@mace registryshadcn components

Infra

SupabasePostgres MigrationsClerk Auth

Languages

TypeScriptC#PythonPLpgSQLShellJavaScript

Impact Metrics

33.65B

TOKENS PROCESSED

95.1%

SERVED FROM CACHE

+1.53M

NET LINES (~20% GENERATED)

138/day

AVG COMMITS

LANGUAGES

296

TOTAL REPOS

Every commit above shipped through fabric — the multi-agent platform we run internally.Explore Fabric →

How We Achieved This→September Development Report→

// Seven months, measured

We publish our own numbers. Every month.

Pulled from the GitHub and Anthropic APIs, not written by marketing. The second chart is the one that changed our business — it counts how many times code reached production, and before June the answer was zero because customers couldn't ship it themselves yet.

Commits

Across every Mace repository

0

2,500

5,000

759

520

1,741

2,493

3,401

2,650

4,146

Mar

Apr

May

Jun

Jul

Aug

Sep

Publishes to production

Times code actually went live

Mar

Apr

May

Jun

Jul

Aug

Sep

Real data — not marketingPulled live · GitHub + Anthropic APIsRead the monthly reports →