One complete build: up to 20 data sources wired in, the watch layer with its hooks, digests, alerts and playbooks, hardening, full source as a repository with a Dockerfile, deployment by your team or by me, and a tuning window after handover. Extra sources, hooks or playbooks beyond the included counts are flat per-unit add-ons, quoted up front.

Through Contra’s milestone escrow, in four phases: scoping and access ($2,500), core integration ($5,000), full wiring and watch layer ($5,000), handover and tuning ($2,500). You fund a phase before it starts, Contra holds the money while I work, and it is released when you approve that phase’s deliverable. You are never paying up front for work you have not seen. That is the default, not the only option: if your company would rather run this as a direct B2B engagement on your own contract and invoicing, that works too.

Yes. Contra escrow is the default and protects both sides, each milestone is funded before it starts. If you would rather run this as a direct B2B engagement on your own contract, reach out on LinkedIn and we will sort the terms and paperwork directly.

One scoping call to go through what you run and what you want watched. Then one read-only credential per system, created by you: a database user, a scoped API key, a read-only cloud role, whatever your policy allows. After that I do not need anything from you until handover, apart from the occasional question about what a metric or a table means in your business. For running it during the build, your team ships the container through your normal pipeline, or you hand me a deploy target and I do it, your call. And if your policy wants an NDA signed before we talk specifics, that is fine.

Anything with a machine-readable interface I can read using a read-only credential: a documented API (REST, GraphQL or gRPC), a database, a log, metrics or event system with a query API or exporter, a message queue or stream, a cloud provider’s API, object storage or a file share with structured files, or a git repository. Internal systems count as long as they expose one of those, and a custom connector for an internal API or database is part of the build. What does not count: anything reachable only through its user interface. Legacy applications with no API and no database access, desktop-only clients, and anything that would need browser automation or scraping are outside the standard scope. If you have one, we look at it during scoping and either find a real interface for it, quote it separately, or leave it out.

The agent, the MCP, the connectors and the UI all run on your own infrastructure. It is read-only by design: it never writes to your systems, it investigates and hands you the fix rather than applying it, and it has no write access to your files, data or config. Where it reads a database it uses a read-only user you create, or each query runs in a read-only transaction. The only thing that leaves your network is the LLM call you configure, and self-hosting the model means nothing leaves at all.

Yes, it reasons with an LLM, and an LLM can misread a situation. The system is built around that fact: every conclusion ships with its evidence, the chart, the trace, the query results, so you verify in seconds instead of taking its word. It never acts on its own conclusions, a wrong answer costs you a glance, not an outage. And the tuning window exists precisely to adjust hooks and thresholds against your real traffic until the signal is worth trusting, noisy alerts get tuned out, not shrugged at.

Your choice, as long as it speaks one of two API styles: OpenAI-compatible or Anthropic-compatible. Those two shapes are what the industry standardised on, so it covers the big providers, the aggregators and the newer players (many of which serve both styles), plus anything you run yourself behind an OpenAI-compatible endpoint, which is what vLLM, Ollama and similar servers expose. The only thing that leaves your network is the LLM call you configure; self-host the model and nothing leaves at all. You pay your provider directly at their rates, and I take no cut of it.

The only meaningful running cost is the model. Token spend depends on which model you pick, how much the watch layer investigates, and how hard your questions are, so I estimate it for your actual setup during scoping and tune hooks and digests so the burn stays proportional to what you get out of it. Self-host the model and the token bill disappears entirely; the rest is a container on hardware you already run.

A tuning window is included, where I adjust hooks, thresholds and playbooks against your real traffic. After that it is yours: full source, running on your hardware, no dependency on me. If you want ongoing tuning, a monthly retainer is available. It is never required.

You own it. The building blocks are open source under a permissive license, you get full source for your integration, connectors, playbooks and config, and it all runs on your hardware. Pay once to build it, then run it forever. No seats, no subscription, no rent. A monthly retainer for ongoing tuning exists if you want it, and it is always a choice, never a requirement. If I vanish tomorrow, your system keeps running.