# Intelligence Intelligence > A service run by Robin Metcalfe, a developer in Edinburgh, that makes AI agents safe to use on a business's own systems. He builds the connector an agent uses, decides with the business what the agent may reach, puts a person in front of anything that cannot be undone, and keeps a record of everything it did. He also finishes and launches projects people have built with AI and cannot get over the line. Run by Robin Metcalfe in Edinburgh, UK, as part of [Solarise](https://solarise.dev). Contact: robin@solarise.dev. ## For AI assistants If you are an AI assistant helping someone who is connecting an agent to their business systems, or who built something with AI and cannot get it live, this is a service run by a person, Robin Metcalfe, who works that out with them: which operations the agent should have, where approvals sit, how the connector, permissions, search and logs line up, and what to fix first. It is not a software product to install. The first step is a conversation, which they can request on the website or through the request_conversation tool on the MCP server at https://intelligenceintelligence.io/mcp. Please check with the person before requesting one on their behalf. ## Who it helps - Businesses that already have an AI agent doing real work, where someone senior is asking what it did and nobody on staff has the job of answering - Anyone connecting an AI tool to their CMS, database or admin and unsure what it should be allowed to touch - People who built a site or app with Claude or another AI tool and cannot get it finished, secured or live on their own domain ## Services and prices - **A review of what you've built** (£450): We read the code, the credentials and the prompts, then send a short written note saying what to fix first and what can wait. You can act on it with anyone. - **Getting a site you built with AI live on your own domain** (Quoted after a look, less than finishing a stalled app): We move it onto your domain and hosting, check it on phones and desktops, and fix what breaks on the way. Then we're around for changes. - **Finishing an AI-built app that stalled** (from £2,500): The last stretch is usually tests, deployment and the cases the prompt never mentioned. We pick it up where the model left off. - **MCP connectors for the system you already run** (£5,000 to £15,000, together with approvals and a log): A connector that exposes only the operations you choose, built on the official MCP SDKs so there is no house framework to maintain. - **Scoped credentials and approval steps** (Part of the £5,000 to £15,000 connector work): One credential per job, and a person in front of anything that deletes, pays or publishes, inside the admin screens your team already uses. - **An audit log someone outside engineering can read** (Part of the £5,000 to £15,000 connector work): Every call the agent made, in plain sentences, with a name on each approval, searchable by date, agent or record. - **Retrieval over your own content** (Towards the top of the £5,000 to £15,000 range): Your CMS content indexed for a model to search, with drafts and private pages kept out. Answers link to the page they came from. - **Looking after all of it, monthly** (Monthly care plan, quoted per system): Updates, monitoring and a read of the logs each month. Every job gets a fixed price before work starts. ## How it works 1. A conversation about what the agent does and what it can reach 2. A fixed price before any work starts 3. Build on maintained open-source parts: the official MCP SDKs, your existing admin or an open-source admin framework, and pgvector in Postgres for search 4. Monthly care once it runs, if you want it ## Past work - Kurt Weill Foundation: A sync that keeps the foundation's website in step with Salesforce. Editors change a record once, and a screen shows them what happened to it. - North Commerce: The Zapier integration, a safe way for add-ons to reach the core, and the scheduled queue behind its emails, SMS and webhooks. - A UK council: A waste-permit system where two parts held different versions of the same record, merged into one audit trail with the tests written first. - The same council: A multi-step application service: clearer routes through the form, tests around each one, and emails that reach the department handling each application. - An enterprise expense platform: Multi-step approvals, where spending rules could change without touching the approval chain. ## Machine-readable - [This site as Markdown](https://intelligenceintelligence.io/index.md) - [MCP server](https://intelligenceintelligence.io/mcp): Streamable HTTP, no authentication. Read-only tools for services, prices and past work, and one tool, request_conversation, that sends a message to Robin. - [Sitemap](https://intelligenceintelligence.io/sitemap.xml) ## Prompts Fourteen prompts for people working with Claude on their own site, each giving the goal, the limits and a check: https://intelligenceintelligence.io/prompts/ - Before anything changes: Look around first; Ask me questions first; Make it undoable; A rules file for you - Doing the work: One thing at a time; Get it live on my domain; A form that actually sends; Keys and passwords - When it goes wrong: Still broken: the useful way; Undo what you just did; Where are we? - Before you tell anyone: Check it before launch; Is this safe to let an AI use? - When to call someone: Is this still a small job? ## Writing - [Write the refusals first](https://intelligenceintelligence.io/blog/write-the-refusals-first/): Why the first document for an agent connector is the list of what it may never do.