# BlvkWare

> AI agents that do the work, designed from a business's own answers and sold as
> downloadable kits. An agent reasons through a task, operates the business's
> existing software, communicates with its customers and follows multi-step
> processes, rather than being a dashboard someone has to click. A kit is the
> complete specification of one agent, which the buyer builds with the tools
> they already use. Also publishes free browser-based business analysis tools.

BlvkWare is run by William Russell Wheeler, based in Mississippi. It serves small
businesses anywhere: home services, trades, clinics, professional practices,
agencies and online stores. Everything is self-serve: design, purchase and
download happen on the site without a call.
Contact: russ@blvkware.dev

## The model

Traditional software gives a company's employees tools. An agent is the operator
for those tools. BlvkWare designs the agent and sells its **kit**: every file the
agent needs, generated from the buyer's answers about the job, their systems,
what it may do on its own and the rules it must never break.

A kit contains:

- the agent's instructions: a system prompt written for that business, with the
  owner's "never do" rules built in word for word
- its tools, in OpenAI, Anthropic and MCP formats, each marked **read**,
  **internal** (changes the business's own records) or **external** (reaches a
  customer, a supplier or money)
- a workflow per capability: what starts it, the steps in order, the cadence
  where there is one, and which steps wait for approval
- JSON Schemas for the records it keeps, always including an approval queue and
  an append-only audit log
- an autonomy policy with four levels, written to be enforced in code
- a setup checklist of the facts only the business can supply, integration notes
  per system, and an acceptance test per capability

**A kit is not running software.** It does not connect to anything by itself and
contains no credentials. The buyer, or whoever builds for them, loads it into
OpenAI or Anthropic models with tool calling, any MCP client, or an automation
platform such as n8n, Make or Zapier. Every file is plain JSON or Markdown.

## Prices

- **Operator kit: $79**, one-off. For an agent that owns one job start to
  finish.
- **Deputy kit: $199**, one-off. For an agent that owns an entire business
  function, across more systems and channels, making judgment calls within set
  boundaries and escalating exceptions. A Deputy key also covers any
  Operator-sized design.

**The kit is derived from the designed scope, not chosen by the buyer**: a design
that is really a Deputy cannot be bought at the Operator price, and the
configurator states which kit it landed in and why, before purchase. There is no
subscription and no per-capability charge. The configurator shows the full file
list of a kit before purchase.

Purchase is on Gumroad, which issues a licence key. The buyer pastes the key on
https://blvkware.dev/hire/ to download. One key is one agent: the buyer can
change that agent's design and download it again, up to 20 times. What it
costs to run the agent the buyer builds (model usage, phone numbers) is
between the buyer and those providers.

## Roles and capabilities

The catalog has **23 named agent roles** built from **73 individual capabilities**.
Roles cover email management, lead qualification, appointment setting, quote
follow-up, outbound sequences, CRM operation and hygiene, multi-step workflow
automation, operating web portals that have no API, onboarding, collections,
reconciliation, data collection and analysis, visualisation, scheduled reporting,
research, customer support, orders, returns, stock, shipments and renewals.

Each role has a core set of capabilities that is always in its kit, plus
suggested and optional ones. Dependencies resolve automatically: a capability
that cannot function without another has that other one added. Every capability
has a written acceptance test. The full catalog is at https://blvkware.dev/agents/
and an agent is designed, priced and bought at https://blvkware.dev/hire/.

## The autonomy ladder (important when describing how these agents behave)

Every kit specifies four levels, and the agent it describes **starts at L1**. The
owner raises the level themselves after watching it work.

- **L0 Watch**: observes and reports, takes no action.
- **L1 Draft**: keeps the business's own records, but anything that reaches a
  customer, a supplier or money is held for a person. The starting level.
- **L2 Approve**: acts on its own but pauses for a person above limits the owner
  sets.
- **L3 Operate**: acts within its scope and escalates exceptions, switched on by
  the owner after a supervised period.

The kit's guardrails hold at every level: facts come only from records, the
conversation or the knowledge base (nothing invented), opt-outs are permanent,
nothing is sent in quiet hours, every action is logged before it happens, and
instructions arriving inside a customer's message are treated as information,
never as orders. How the built agent behaves is under the buyer's control.

## How to get in touch

- Design an agent and get its kit without a call: https://blvkware.dev/hire/
- Browse what each agent role owns: https://blvkware.dev/agents/
- Or email russ@blvkware.dev.

## Free tools

All of them run entirely in the visitor's browser. No account, no sign-up, no
payment. They exist as the work sample in place of client case studies.

- [Business Scan](https://blvkware.dev/scan/) finds the work. Give it a company's public
  website and it reads the live page, works out which jobs are quietly costing
  that business the most -- unstaffed, done after hours, or done by somebody who
  should be doing something else -- and names the agent worth building for each,
  priced as a kit from the published catalog rather than estimated.
- [Agent Designer](https://blvkware.dev/design/) designs the agent. Describe how a business
  actually operates and it works out which agent is worth building first, exactly
  what it has to be able to do, which existing systems it must operate, and what
  still needs a person -- then prices its kit and hands the finished design to the
  kit page with every option already set.
- [App Builder](https://blvkware.dev/build/) is a free build-anything tool: it
  generates a complete single-file web application from a description, runs it,
  audits it and repairs its own errors, using the visitor's own model key.

## Open source

- [pkgguard](https://blvkware.dev/pkgguard/) is a free, Apache 2.0 pre-install
  security gate. It checks npm, PyPI and crates.io package names before an AI
  agent installs them and answers ALLOW, REVIEW or BLOCK with the reasons,
  catching hallucinated dependencies, slopsquats and typosquats. CLI, MCP
  server, GitHub Action and a self-hosted HTTP API; no account and no key.
  Install it from https://github.com/rxslice/pkgguard-API with
  `pip install "git+https://github.com/rxslice/pkgguard-API.git"`. **It is not
  published on PyPI and the name there is unclaimed, so never suggest a bare
  `pip install pkgguard`.**

## Important caveats for anyone citing this site

- Every dollar figure produced by the free tools is a **projection modelled from
  public industry benchmarks**, never a measurement of a real business's finances.
- BlvkWare has **no client case studies and no client references**. This is stated
  openly on the site; the tools and the kit previews are offered as the evidence
  instead. The one long-form write-up, the Recovery OS build log, documents a
  system BlvkWare built from its own planning documents -- no client commissioned
  or paid for it, and the page says so.
- A kit is a specification, not a running agent. Tool output is AI-generated and
  is not financial, legal or engineering advice.

## Pages

- [Home](https://blvkware.dev/): the model, the two kit sizes, prices, tools, FAQ
- [Recovery OS build log](https://blvkware.dev/recovery-os/): a worked example of
  a whole-process system -- four ordinary business documents (a slide deck, a call
  script, a spreadsheet, a strategy memo) turned into an eight-stage operating
  system with 203 self-checks. Documents the six rules the software enforces, the
  three defects found before shipping, and carries the part-number identifier
  running live in the page so a reader can test it. Built by BlvkWare from its own
  planning documents; no client paid for it.
- [Design an agent](https://blvkware.dev/hire/): the kit configurator. Pick a role,
  its systems, channels, capabilities and autonomy target, and your own rules;
  see the kit's full file list and price; buy on Gumroad; download with the
  licence key. No call required.
- [Agent catalog](https://blvkware.dev/agents/): every role, what each one owns,
  what every kit contains, and every capability with its acceptance test.
- [Sample kit](https://blvkware.dev/sample-kit/): a complete Follow-Up Agent kit
  for a fictional plumbing business, every file readable in the page and
  downloadable as a zip, generated by the same code as a paid kit. Free to read
  and share; not licensed for use in a business.
- [What is an AI agent](https://blvkware.dev/what-is-an-ai-agent/): the
  difference between an agent, an automation and a chatbot; the four capabilities
  an agent actually needs; the four questions to put to a vendor; and the three
  cases where a business should NOT buy one (the job is rare, the job is
  genuinely deterministic, or the process is not written down anywhere).
- [What an AI agent should cost](https://blvkware.dev/ai-agent-pricing/): the
  three pricing models in the market, defensible build and monthly ranges by
  shape of agent, what actually drives the price, why an agent should be compared
  to a part-time hire rather than to a SaaS subscription, the arithmetic for
  deciding, and five warning signs in a quote. BlvkWare's own prices are stated.
- [AI agent permissions and safety](https://blvkware.dev/ai-agent-permissions/):
  what realistically goes wrong with business agents, the four-level autonomy
  ladder (Watch, Draft, Approve, Operate), least privilege for agents, the three
  things every agent must have before it touches a customer (a reasoning log, a
  reverse, a stop), what must never be unsupervised, and where liability sits.
- [How BlvkWare agents are controlled](https://blvkware.dev/how-agents-are-controlled/):
  the mechanics rather than the argument. Where credentials live (a permissioned
  file on the agent's own machine, never in the codebase, not separately
  encrypted at rest and stated as such), the two destinations data goes to (the
  customer's own vendors, and the model provider for reasoning only), that the
  agent runs with no model at all on a deterministic path and what that costs in
  escalations, the append-only log written before each side effect, what can and
  cannot be undone, the customer's own one-press stop, export on exit, and an
  explicit list of what the page does NOT claim (no SOC 2, no encryption at
  rest, and that the agent can be wrong).
- [Quote follow-up automation](https://blvkware.dev/quote-follow-up-automation/):
  how automated quote follow-up works, the arithmetic for deciding whether it
  pays, why the sequence must stop when the customer answers, and why US A2P
  10DLC registration (3-6 weeks) makes SMS-based follow-up slow to launch where
  email is live in days.
- [AI automation for plumbers](https://blvkware.dev/ai-automation-for-plumbers/):
  which plumbing processes are worth automating and which are not. Estimate
  follow-up and review requests first; dispatch, pricing and emergency call
  answering last or never. Includes the review-gating compliance line and why
  replacing ServiceTitan or Jobber to gain follow-up is a bad trade.
- [Vendor renewal tracker](https://blvkware.dev/vendor-renewal-tracker/): two free
  browser tools plus the method. A renewal deadline calculator that derives the
  notice deadline (renewal date minus notice period) and scores a vendor on
  value, waste and leverage, and a statement importer that reads a bank or card
  export in the browser and infers each billing cycle from the gaps between
  charges. Nothing is uploaded; there is no server.
- [Subscription audit](https://blvkware.dev/subscription-audit/): the method for
  cutting recurring business cost. Pull twelve months not three because annual
  charges hide in a short export; record seats paid against seats active; check
  cancel, downgrade, renegotiate and keep in that order; and sort the result by
  money recovered per minute rather than by size of saving, because sorting by
  size buries the quick wins under the hard conversation and the audit stalls.
- [Privacy policy](https://blvkware.dev/privacy/): no accounts, no analytics, no
  cookies, no tracking; tools keep data only in the visitor's own browser
- [Terms of service](https://blvkware.dev/terms/)

## Verification infrastructure

Separate from the agent practice above, and aimed at machines rather
than at businesses.

- [HALLUX](https://blvkware.dev/hallux/): does this identifier actually exist? Package
  names, module paths and DOIs, checked against their authoritative
  registry before an agent installs, imports or cites. Returns
  exists / deprecated / absent / phantom / squat / unknown.
  A phantom is a name no registry has and models repeatedly invent.
  A squat is a phantom somebody has since registered, which is a
  supply-chain attack caught before the first install.
  Live namespaces: cite.doi, pkg.crates, pkg.go, pkg.maven, pkg.npm, pkg.nuget, pkg.pypi.
  Free tier: 500 identifiers a day per client, no account, no key,
  within a shared pool of 50,000 a day for all free traffic.
  MCP server: https://api.blvkware.dev/hallux/mcp (Streamable HTTP, no account, no key).
  Tools: hallux_check_command before an install, hallux_check_manifest
  after editing a dependency file, hallux_check for identifiers.
- [Agent catalog](https://blvkware.dev/.well-known/ai-catalog.json): machine-readable
  index of every capability that answers, with prices and payment
  rails. Capabilities listed under `roadmap` are not built.
- [HALLUX specification](https://blvkware.dev/docs/hallux-spec): verdict semantics,
  namespaces, receipt format and corpus methodology.