An agent that reads your documents and systems, works out the next step, and shows it to a person before anything is written. That one boundary is what separates an agent you can run in production from a demo that looked good once.
None of this is theoretical. Each pattern below links to the live system or the write-up of the decisions behind it.
The model never writes to your records directly. Every action it wants to take is rendered as a dry run a person approves or dismisses. A wrong answer costs one dismissed card, not a corrupted ledger. This is how Sable handles money.
How Sable keeps the model away from money →The agent answers only from passages it retrieved from your documents, cites them, and declines when nothing relevant clears the similarity threshold. The assistant on this site runs exactly this way on a Cloudflare Worker.
Try the live agent on this site →Agents get governed access to your systems through Model Context Protocol servers. Read access ships first, write access only behind the confirm step, and every call is logged. This site publishes its own MCP server you can connect to from Claude.
The operating model, with a 30-day plan →Where privacy is the requirement, the agent runs against on-device SQLite with no cloud backend at all. Privacy becomes a property of the architecture rather than a promise in a policy.
Sable, the local-first finance agent →Answer from your documents on your website or WhatsApp, hand off to a person with the full context, and draft replies a human sends.
Read invoices, tickets or forms, propose the next action in your CRM or ERP, and wait for the click that commits it.
Pull from your database and documents on a schedule and produce a briefing, the way Sable's morning briefing works, with every number traceable.
An MCP server over your Postgres, Git or admin tools so your team can ask questions and run approved actions from inside the AI they already use.
A document-answering agent on its own is priced on the AI chatbot page, from $800. Wiring agents into systems you already run is scoped on the business services page.
Which systems the agent reads, which actions it may propose, and which of those, if any, it may take without review.
Retrieval and MCP read access ship first, so you can watch the agent reason over real data before it can change anything.
Write actions arrive as review cards. The boundary lives in code and is logged, so it holds when the model is wrong.
A test set of real questions and actions, run on every change, and the repository in your hands from the first commit.
It is quoted as a fixed price after a short scoping call, because the cost depends almost entirely on how many systems the agent has to read and write and how much of that needs the confirm step. Smaller pieces of work are $15 to $30 an hour. A document-answering agent on its own starts at $800 on the AI chatbot page.
Only the ones you have approved it to take without review, and the default is none. Everything else is proposed and waits for a person. That boundary is in the code, not in a prompt, so it holds even when the model is wrong.
Whichever fit the job and your constraints. The agents on this site run on Cloudflare Workers with Workers AI and Vectorize; Sable runs a model against local SQLite with no backend. Claude, OpenAI and local models through Ollama are all options, and the architecture does not lock you to one.
Yes. An MCP server exposes your data and approved actions as tools, and your team connects it from Claude or any MCP-capable client. This site's own MCP server is live and documented, so you can see the pattern before paying for one.
Model usage and hosting, estimated before work starts, because that is what surprises people later. The edge stack used here has a generous free tier; a local-first agent has almost none.
You do. The code is in your repository from the first commit and runs on infrastructure you control, so you can maintain it yourself or hire anyone else to.
Tell me what the agent should read, what it should be allowed to propose, and which systems it touches. You'll get a fixed price and an estimate of the monthly running cost.