The Accelerator
Infrastructure for software that runs at AI speed
Model-agnostic generators do the building. Independent judges verify the work. Interlocks limit what can reach production. Your operators steer the whole thing.
Architecture
Five layers. The models are the replaceable part.
Why not just use an AI model?
Claude is a brilliant contractor. The Accelerator is the team that remembers.
On its own, a model has no memory of your business, no hands on your accounts and no safety rails. Every client site gets its own crew of specialist agents and its own private Library, which grows smarter with every change.
Your Library: the memory of your site
Every decision, design, conversation, change and lesson, organized by your site and private to you. Ask it why something works the way it does, and it answers with the decision and the conversation where it was made.
Three tiers of know-how
Next.js upgrade playbook · database security rules
Brand voice · pricing rules · tone with customers
Catalog rules · checkout quirks · what must never break
In your own repository
Everything needed to understand your site without us: the handbook, decisions, approved designs, site skills and tests, as plain Markdown and code.
In your private Library with us
How the work happened: conversations, evidence, previews, the tool log and the release timeline. Exportable in one click, any time.
Your crew knows 47 things about Machine Shop
- DecisionThe assistant only mentions products its search returned · from Work #4
- ReleaseSearch by brand, live with before/after screenshots · Work #7 · undo
- Site skillCatalog is read-only; every query is paginated · v2 · approved
- DesignSaved lists for shoppers · approved · in build
- LearnedMobile search box too small; crew proposes a layout rule · awaiting you
Machine Shop is our test site. See its real Library in a walkthrough, not on slides.
The Accelerator
Everything it takes to hand over the keys
Application Model
Your architecture, invariants, money and identity surfaces, and what “done” means — written down, versioned, and checked on every change. The machine reads it; so can you.
Independent Assurance
The model that writes the change never grades it. Held-out tests, invariant checks, historical replay, synthetic users, and preview telemetry judge every change independently.
Interlocks
No production write keys in an AI session. Money, auth, and destructive migrations never auto-promote. Spend caps, loop limits, and full provenance on every change.
The Discipline Pack
The operating habits that made Customer Zero work — a rulebook, a numbered backlog, design-before-code, every incident becomes a regression test — installed on your app.
Mission
What the application should accomplish — and the bounds it may never cross.
Work
What the Accelerator is doing right now, queued, held, or done.
Health
Is the application behaving correctly — tests, invariants, and live telemetry.
Changes
Every change, in plain language, with the intent that caused it.
Confidence
Why the Accelerator believes each change is safe — the evidence package.
Controls
Raise or lower autonomy, hold a change, roll one back, pin a model.
Evidence, not code review
The builder never grades its own work.
An AI that writes both the code and its tests can confidently prove its own misunderstanding. So the Accelerator separates the actor from the judge. Every change arrives with an evidence package that answers one question: why is this safe?
- Held-out tests the builder can't edit
- Business invariants, checked on every change
- Replay against historical behavior
- Synthetic users on a test database
- Preview telemetry before release
- People judge outcomes, not diffs
Dashboard balances must match every folio
- Root cause
- Unordered paged read repeated/skipped rows
- Surfaces touched
- Dashboard, integrity checks, directory
- Regression test added
- 1,280 ledger rows · concurrent reads
- Full suite
- All tests passed
- Invariant
- Balance = Σ posted ledger rows
- Invariant
- One pricing function, every surface
- Preview deploy
- Staging data reconciled to the penny
- Promotion
- Money class → operator approval
Modelled on a real Customer Zero fix: dashboard balances that disagreed with the ledger.
The Toolbelt · secured by design
Your keys never enter a prompt.
Your crew reaches GitHub, Vercel and your database through one governed gateway, built on the open Model Context Protocol (MCP). It is the same pattern large engineering organizations use to let agents act on real systems safely, sized for one website.
Agents never hold your keys
Your credentials sit in an encrypted vault. Agents call named tools, and the Toolbelt attaches the credential on our server. No key ever appears in a prompt, a chat or a log.
Connect, don't paste
GitHub, Vercel and your database connect with sign-in buttons. Each connection is tested at setup, so problems surface on day one, not at 11pm.
Short-lived, single-job passes
Every job gets a pass that works for one site, one task and one specialist, then expires. Your tenant comes from the pass, never from anything an agent types.
Each specialist sees only its tools
The builder can't see the release tool. The judge can't edit code. Least privilege is enforced by what exists in the agent's world, not by asking it nicely.
Gates live inside the tool
Going live, data changes and anything touching money or sign-in check the gate policy at the moment of the call. An agent can't talk its way past a gate.
Every action on the record
Each call is logged with who, what, when and what it cost. That record becomes the evidence on every change and the timeline in your Library.
- 1Agent asks for a tool, with a single-job pass
- 2Toolbelt finds your site from the pass
- 3Checks the specialist may use this tool
- 4Checks the gate (money, data, sign-in, go live)
- 5Attaches your credential from the vault
- 6Calls the service, hides sensitive data, logs it
Autonomous coverage
You set the ceiling. The evidence raises it.
Autonomy is earned one class of change at a time. The measure is simple: what share of your application's changes can the Accelerator safely own?
Typical commissioning exit: AC1 live, AC2 on selected work
The shift
Software was never a factory.
A factory moves a part down a line, with hands at every station. Software never worked that way: it circles through build, test, refactor and ship. AI makes each turn faster than people can inspect. So stop inspecting parts and run the machine.
The software factory
Linear. Human hands at every station. Throughput limited by review.
- Ticket
- Developer
- Pull request
- Code review
- QA
- Deploy
The accelerator
A loop that gets faster each turn. Machines build, independent judges verify, and people steer from the control room.
Commissioning cohort 01 · now forming
Put your application under power.
We're commissioning a small number of production web apps next. Tell us where yours lives. We'll reply with a scoping call and an honest read on fit.