# Your company has the answers. Now you can find them. Connect data, documents and decisions. Middleout brings your company's knowledge together to answer questions with evidence and remember what it learned. ## Your numbers don't add up. And nobody remembers why. Finance and Sales report different net sales for the same month. The explanation lives in a spreadsheet, in someone's head or in a meeting that already happened, so every month the same question gets asked again. The important questions get asked in one meeting and forgotten by the next. ## You ask, and your company learns. Every answer is ready for the next question. ### Connect your sources Middleout studies every table and spreadsheet. It uses their descriptions when they have them; when they don't, it works out what each field means from its name, and asks you about what it can't figure out. ### Ask and investigate Ask in your own words. Middleout checks several possible causes in your data, rules out the ones that don't fit and shows you the answer with its numbers. ### What you learn, stays Every answer is saved next to your metrics and decisions. Your team can correct or complete it, and those corrections count for more than what the AI assumes. ### Share it with your team Ask for a dashboard or a presentation in one sentence. If a number changes later, it updates by itself and Middleout tells you. And you share the findings on Slack or Google Chat. ### Use it from Claude or ChatGPT too If your team already uses Claude, ChatGPT, Codex or Cursor, connect them to Middleout and they'll answer with your company's data and definitions. Each person sees only what they're allowed to see. ## It's not one more chat over your data. With Middleout, your company remembers what it learns. ### Certified definitions Your team approves each definition, and Middleout answers only with certified ones and tells you which it used. ### It resolves contradictions When two teams calculate differently, it shows you the gap and asks which one counts. The agreement holds for everyone. ### It remembers decisions Every decision and conclusion is kept with its evidence. The next question starts from there. ### You build it in one sentence Ask for the dashboard or the presentation in your own words, and it keeps itself up to date. ## The frontier of what your company knows. Every answer pushes it a little further. Beyond what your company knows are the hard questions: where to grow, which customers are leaving. Middleout investigates them, and every accepted answer pushes that edge out. ## Your data stays yours. Middleout uses it to answer, and doesn't keep it. We know what it's like to walk into the committee with a number that doesn't match Finance's. That's why Middleout shows where every figure comes from and doesn't keep your data. ### Your data doesn't stay in Middleout Databases like BigQuery are queried where they live. To query a spreadsheet, Middleout uses a temporary copy that it deletes shortly after. ### Read-only Middleout can't modify your sources: it reads BigQuery, Drive and Sheets with read-only permissions. ### Each person sees their own Each person sees only what their permissions allow. An assistant like Claude or ChatGPT, connected to Middleout, sees the same as the person who connected it. ## Stop waiting weeks for an answer. Ask and get the answer in minutes. When a number changes, your dashboards and presentations update themselves. 1. **Request early access**: We'll let you know when we open for your company. 2. **Connect your sources**: Read-only. Middleout studies them and asks you about what it can't figure out. 3. **Ask, investigate and share**: Whatever you build updates every time a number changes. Your data stays in your systems. Each person sees only what they can already see. ## Frequently asked questions ### What does Middleout keep about my company? What it takes to understand your data: what each metric means in each team, what was decided and why, and how it all relates. It learns this by reading your data, documents and code, and it grows with every question it answers. It doesn't keep your tables or your files. ### Which sources can I connect? BigQuery and Google Sheets to query data; Google Drive for documents, presentations and meeting recordings; and GitHub for repositories and their documentation. ### Can Middleout modify my data? No. Middleout is read-only: it can't modify your data or your sources. ### Do I need a ChatGPT or Claude subscription? No. Middleout has its own app for asking questions, investigating, and building dashboards and presentations, and AI is included. Connecting it to an assistant is optional: it's useful if your team already works with one. --- # You ask the questions. Middleout does the investigation. Middleout investigates the way a good analyst would: it considers several possible causes, looks in your data for what confirms or rules out each one and reaches a conclusion with its numbers. You decide what's worth investigating. ## How an investigation works. ### It starts with a business question An investigation is born from a question that matters: why repeat purchases dropped, why one store sells less. You open it from the chat or ask Middleout for it directly. It has an owner and starts as yours; when it's ready, you publish it to your team or the whole organization. ### Several causes at once A rushed analysis settles for the first explanation that fits. Middleout proposes several possible causes at the same time and puts them up against each other. You can add the one you suspect: your intuition enters the investigation as one more cause and is checked with the same rigor. ### Every number, for or against Middleout searches your sources for the numbers that confirm or rule out each cause and notes where they come from: which table and which definition they were calculated with. What has no data yet isn't ruled out: it stays open, in plain sight, until there's something to answer it with. ### A conclusion your team accepts When a cause is confirmed, Middleout proposes a conclusion tied to the number that makes it true. Nothing becomes part of what your company knows until a person accepts it. When you accept it you can note what was decided, and that decision is recorded next to its evidence. ### It keeps being checked A conclusion isn't filed away and forgotten. Middleout checks it again against each day's data and tells you if it stops holding. The new questions it left open are ready for the next investigation. ### It stays in memory The whole investigation, with what was ruled out and why, stays in your company's memory. The next time someone asks something similar, Middleout starts from there instead of from zero. And from the investigation you ask, in one sentence, for the dashboard to follow the number or the presentation for the committee. ## Questions about this ### How does what my company knows grow? Through its people. Every time someone asks, analyzes or investigates with the company's data and documents, they find out something nobody knew before. Middleout keeps it so it isn't lost, and what your company knows grows with every question. ### How do I get started? Request early access. When we open for your company, you connect your sources and start asking and investigating. --- # Everything your company knows, connected. And available to the AI you already use. AI isn't short on intelligence: it doesn't know your company. Middleout reads your data, documents and code and builds your company's memory: what each metric means, what was decided and how it all relates. Every answer, in Middleout or in your assistant, uses it and makes it grow. ### Why your numbers don't match Each team calculates its own way. Middleout keeps both definitions, says which one it used, and your team agrees on the one that counts for the whole company. ### Your assistant, with your company's context Connect Middleout once to Claude, ChatGPT, Codex or Cursor and ask from there. Your assistant only sees what you can see. ## How your company learns. ### It learns from your sources When you connect a database or a spreadsheet, Middleout studies every table. If the columns have descriptions, it uses them. If not, it works out what each field means from its name. What it can't understand with confidence it doesn't make up: it asks you, so someone who knows the business can fill it in. ### What it keeps The memory keeps what gives your data meaning: metrics with their formula and the tables they come from, agreed definitions, decisions with their alternatives and assumptions, rules, procedures and learnings. It also keeps how your people explain the causes of what happens in the business. All connected: every piece knows what else it relates to. ### Definitions are agreed on When two teams or two sources say different things about the same thing, Middleout doesn't pick one on its own. It shows both versions and a person decides which one holds for the whole company. From then on, answers use the agreed definition. ### Your people teach it Anyone can teach Middleout a definition, a rule or a decision, from the app or from their assistant. What is only yours counts right away; what is for a team or the whole company waits for someone responsible to accept it. What people confirm weighs more than what Middleout works out. ### It grows with use Every accepted investigation, every recorded decision and every connected meeting adds to the memory. That's why the next question is answered better than the last: it starts from everything your company has already learned. ### Each person sees their own A source can be personal, visible only to whoever connected it, or the organization's, shared with everyone. Each person, and each assistant connected in their name, sees only what their permissions allow. ## Questions about this ### What does Middleout keep about my company? What it takes to understand your data: what each metric means in each team, what was decided and why, and how it all relates. It learns this by reading your data, documents and code, and it grows with every question it answers. It doesn't keep your tables or your files. ### Which sources can I connect? BigQuery and Google Sheets to query data; Google Drive for documents, presentations and meeting recordings; and GitHub for repositories and their documentation. --- # Three ways to take an answer to your team. All three start from the same question. When you ask, Middleout builds the investigation with its evidence. From there you ask, in one sentence, for the dashboard to follow the number or the presentation to tell it to the committee. ### Investigation The question, the causes that were checked, the numbers that confirm or rule them out and the conclusion, in one place. Your team reviews and accepts it. ### Dashboard The numbers you follow every week, each with its query and definition in view. ### Presentation The slides for the committee, with figures that point to their source instead of copying it. ## Same question, three ways to carry the answer. In Middleout, a question can end up as an investigation, a dashboard or a presentation. All three start from the same question and point to the same numbers, so what one shows, the others show too. ### Investigation It's where a question gets answered with a method. Middleout proposes several possible causes at once, searches your sources for the evidence for or against each one and rules out the ones the data contradicts. What has no evidence yet stays open, not ruled out. The conclusion is tied to the number that makes it true. Your team reviews it and accepts it, and Middleout checks it again against each day's data. If it stops holding, Middleout lets you know. Once you accept it, the investigation is saved to your company's memory with its question, its evidence and its conclusion, ready for the next question. Use it when you want to understand why something happened. ### Dashboard It's where you follow the numbers that matter. You ask for it in one sentence and Middleout builds each chart with its query and the definition it used, in view. You can set a limit on any number and Middleout tells you when it's crossed. The data downloads as CSV or XLSX, or goes to Google Sheets. Use it when you want to know how something is going, week by week. ### Presentation It's where you tell a conclusion to whoever has to decide. You ask for it in one sentence and Middleout builds the slides from the investigation or the dashboard. Every figure on a slide points to its number instead of copying it. When you present it, that version is kept exactly as it was shown. Use it when you have to take an answer to a committee. ## If a number changes, it changes in all three. And Middleout lets you know. The investigation, the dashboard and the presentation point to the same number. If a late return comes in, the number is corrected in all three and nobody has to redo the slides by hand. ### Everything you can build - **Chat with your sources**: Ask about your data and documents in plain language. - **Dashboards you ask for in a sentence**: Ask for a chart and Middleout builds it, query included. - **Presentations you ask for in a sentence**: Slides point to their numbers instead of copying them. - **Exportable reports**: Download them as CSV or XLSX, or send them to Google Sheets. - **Alerts**: Set a threshold and Middleout lets you know when it's crossed. - **Collaborative investigations**: Questions, causes and conclusions, worked on as a team. ## Questions about this ### What happens to my presentations when the data changes? Every slide, widget and conclusion points to its number instead of copying it. If the data changes, it updates in all three and Middleout tells you what changed and why. Anything already presented is kept as a published version. ### How do I get started? Request early access. When we open for your company, you connect your sources and start asking and investigating. --- # Your assistant, with your company's context Add Middleout as a connector, authorize with your account and ask as usual. Your assistant answers with your company's definitions, decisions and data, and only sees what you can see. 1. Add Middleout as a custom connector, with the address of its MCP server. 2. Authorize with your account. The assistant will see only what you can see. 3. Ask as usual: the assistant answers with your company's definitions, decisions and data. ## What your assistant does with Middleout. ### It answers with your company's context Your assistant asks Middleout before answering. So when you write "how did we do in August?", it answers with your numbers, with the definition of sales your company agreed on and saying where each figure comes from, instead of guessing. ### It searches what your company already knows It can search your company's memory for metrics, definitions, decisions and learnings, read an investigation with what the team already concluded and see what each thing relates to. ### It records what comes up in the conversation If a decision is made in a conversation, a definition is clarified or a question is left unanswered, your assistant can record it in Middleout in your name and with its origin. What is only yours stays right away; what is for your team or the company waits for someone responsible to accept it. ### For people who write code If you use Claude Code, Codex or Cursor, your assistant can also read the instructions of connected repositories and the business definitions while it writes code. That way, the code that calculates a metric can use the same definition as the rest of the company. ### What it can't do Your assistant sees exactly what you can see in Middleout, nothing more. It can't modify your data sources. You disconnect it whenever you want from Settings. ### Which assistants it works with Claude, ChatGPT, Codex, Claude Code, Cursor and any other that supports MCP, the open standard assistants use to connect to other tools. You connect it once, with your account. ### You don't need Claude or ChatGPT Middleout has its own app for asking questions, investigating, and building dashboards and presentations. Connecting it to an assistant is optional: it's useful if your team already works with one. ## Questions about this ### How do I connect it to ChatGPT, Claude, Codex or Cursor? You add Middleout as a connector in your assistant, sign in with your Google account and you're done: your assistant answers with your company's data and definitions. It works with any assistant that supports MCP, the standard Claude, ChatGPT, Codex and Cursor use to connect to other tools. ### Do I need a ChatGPT or Claude subscription? No. Middleout has its own app for asking questions, investigating, and building dashboards and presentations, and AI is included. Connecting it to an assistant is optional: it's useful if your team already works with one. --- # Your data stays yours. Middleout uses it to answer, and doesn't keep it. Middleout queries your data to answer and doesn't keep a copy. What it keeps is what it needs to understand it: what each column means, how each metric is calculated and what your team decided. ## Your data doesn't stay in Middleout Databases like BigQuery are queried where they live. To query a spreadsheet, Middleout uses a temporary copy that it deletes shortly after. ## We keep the context What Middleout keeps is what gives your data meaning: which tables exist, what each column means, how each metric is calculated and what your team decided. ## Read-only Middleout can't modify your sources: it reads BigQuery, Drive and Sheets with read-only permissions. ## Each person sees their own Each person sees only what their permissions allow. An assistant like Claude or ChatGPT, connected to Middleout, sees the same as the person who connected it. ## Every number shows its source Every answer says which definition it used, which tables it queried and how recent the data is. If something doesn't add up, you can see where. ## Disconnect whenever you want You remove Middleout's access from your Google or GitHub account, and an assistant's from Settings. ## Questions about this ### Can Middleout modify my data? No. Middleout is read-only: it can't modify your data or your sources. ### Which sources can I connect? BigQuery and Google Sheets to query data; Google Drive for documents, presentations and meeting recordings; and GitHub for repositories and their documentation. --- # The frontier of what your company knows. Every answer pushes it a little further. Beyond what your company knows are the hard questions: where to grow, which customers are leaving. Middleout investigates them, and every accepted answer pushes that edge out. ## An idea that comes from science. ### In science In science, the frontier of knowledge is the state of the art: the furthest anyone has got on a subject. What lies inside can be taught, cited and used. Outside are the open questions, the ones nobody can answer yet. Research is work done right at that edge: starting from everything that is known to answer something nobody had answered. ### Every answer opens new questions When an investigation closes, the frontier moves a little outward, and new questions almost always appear that couldn't even be asked before. That's why the frontier never ends: it grows with every answer. ### The same happens in a company Your company has a state of the art too: its agreed definitions, its metrics, its decisions and what it has already investigated. Outside are the questions that shape its future and that nobody has managed to answer with data: where to grow, which customers are leaving and why, what price the market will bear, which category to open. In most companies that frontier is invisible: nobody quite knows what is known and what isn't, and those questions get settled by opinion. ### Not every question is at the frontier Why margin dropped in August or how much a store sold are important questions, but they're answered from the inside: with the data and definitions the company already has. Frontier questions are a different kind. They're strategic, they cut across several teams and their answer isn't stored anywhere: it has to be built. ### What isn't known, in plain sight Middleout makes the frontier visible. What is backed by data goes into your company's memory, with its numbers and where they come from. What has no answer yet stays as an open question, in plain sight, instead of getting lost in a meeting. ### How Middleout moves it Frontier questions aren't answered with a query: they're investigated. Middleout starts from everything your company already knows, proposes several possible answers, looks in your data for what supports or rules out each one, and keeps in plain sight what has no backing yet. When your team accepts a conclusion, it becomes part of your company's state of the art, and the next hard question starts from there. ## Questions about this ### How does what my company knows grow? Through its people. Every time someone asks, analyzes or investigates with the company's data and documents, they find out something nobody knew before. Middleout keeps it so it isn't lost, and what your company knows grows with every question. --- # We build products that make companies more intelligent. Nexus Labs was born inside companies. Its founders worked within organizations and saw, again and again, how hard it is for a company to know what it already knows. Today we build the products we wish we had back then. ## We saw the problem from the inside ### We worked within organizations Before Nexus Labs, the founders worked inside companies, close to their data and to the people who make decisions. There we saw that what an organization knows lives scattered across people, spreadsheets, systems and conversations, and that pulling it together to answer one question can take weeks. ### The same gaps everywhere One team doesn't know what another already solved. Each area has its own version of the numbers. Customers are known in pieces. And when someone leaves, what they learned leaves with them. ### Today those gaps can be closed What used to take months of manual work can now be solved by combining three disciplines: data engineering to connect and organize the sources, data science to find patterns and explain causes, and artificial intelligence so anyone can ask in their own words and nothing learned gets lost. ## Helping what a company knows grow faster Our purpose is to create products that speed up the growth of an organization's internal knowledge. It isn't about having more data, but about understanding better: ### Its customers Who they are, what they need and why they buy or leave, with the evidence in sight. ### How it works inside Which processes hold the business up, where they get stuck and what changes when a decision is made. ### What other teams know Less distance between areas: what one team discovers is available to everyone, with the same definitions. ## What we believe ### What is known can be traced Every conclusion shows the evidence behind it. ### What isn't known is said out loud Open questions stay in plain sight, not hidden. ### Knowledge adds up Every answer is the starting point for the next question. ### Anyone can ask What a company knows grows with the work of all its people, not only a data team. ## Middleout Middleout brings a company's data and documents together with AI to answer what nobody knows today, and keeps every answer for the next question. ## From Chile, for companies anywhere Nexus Labs is a Chilean company that builds for companies in any country. --- # What is Middleout > Your company's knowledge layer, available in Middleout and in the AI you already use. Middleout is a company's knowledge layer. It brings what each area knows into a shared memory: its metrics, its definitions, its decisions and the data sources where the figures live. That memory can be queried from Middleout or from assistants such as Claude or ChatGPT. Middleout is a Nexus Labs product. ## The problem it solves In most companies knowledge is scattered: each area has its own dashboards, its own spreadsheets and its own definitions of the same words. "Net sales" means one thing in finance and another in sales, and the reason a decision was made lives in someone's head or in a chat thread. There is no shortage of dashboards. What is missing is connected knowledge. ## How it works 1. **You connect your sources.** Databases, spreadsheets, Drive folders, GitHub repositories and meeting recordings. Middleout only reads: it never modifies your data. See [Connect your sources](/docs/connect-sources). 2. **Middleout understands.** It studies each source and builds its **memory**: concepts with an agreed definition, metrics, decisions, rules and how they relate to each other, in a knowledge graph. 3. **You work with that knowledge.** You ask in the chat and the result becomes **artifacts** (investigations, dashboards and presentations) that update themselves when a figure changes. 4. **The memory grows with use.** What is learned in each conversation, investigation or meeting stays in the memory, and people's corrections weigh more than inferences. See [Concepts](/docs/concepts) for the details of each piece. ## Middleout in the AI you already use Middleout's memory is available over **MCP** (Model Context Protocol), the standard AI assistants use to connect to external sources. When you connect Middleout to Claude, ChatGPT, Cursor or another compatible client, the assistant answers with your company's knowledge and sees exactly what you can see. See [Connect Middleout over MCP](/docs/mcp). ## Who it is for For companies whose knowledge is spread across teams and tools that do not talk to each other. Usually the data team sets it up (connects the sources and provides the general context) and the business areas use it to answer questions, track their metrics and make decisions. --- # Concepts > The pieces of Middleout and how they relate: memory, data, artifacts, alerts and roles. ## Memory The memory is what Middleout knows about your business and what you can still teach it. It is organized as a knowledge graph: each piece is an object, and objects relate to each other. The main object types are: | Type | What it is | | --- | --- | | Metric | A number the business tracks, with its formula and the tables it comes from. | | Definition | A business word with an agreed definition. | | Decision | Something someone decided: the question, what was chosen, the alternatives and the assumptions. | | Rule | An agreed way of doing or interpreting something. | | Procedure | Steps that repeat. | | Learning | Something discovered while working with the data. | | Mental model | How someone explains the causes of something that happens in the business. | | Table | A table from a data source, with the description of its columns. | | Document | A document from a source, such as a Drive file or a file in a repository. | ### Concepts, graph, corrections and proposals - **Concepts**: the business words with an agreed definition. - **Graph**: every concept Middleout knows and how they relate. - **Corrections**: when two people or two sources say different things about the same thing, Middleout does not pick one on its own. It shows both versions and a person decides which one holds. - **Proposals**: relations someone proposed, which can be confirmed or dismissed. What is confirmed weighs more when Middleout reasons. ## Data The sources you connect: databases, spreadsheets, folders, repositories and meetings. Middleout studies them and what it understands goes into the memory. See [Connect your sources](/docs/connect-sources). A source can be **personal** (only the person who connected it sees and queries it) or **organization** (shared, and also used by the work Middleout does on its own, such as alerts and scheduled updates). ## Artifacts What you produce working with Middleout. All three are connected to the figures they come from. - **Investigation**: keeps a business question alive. It has an owner, hypotheses, evidence, the dashboards where it is tracked and a conclusion. It starts as something personal and can be published to the team or to the organization. - **Dashboard**: you ask and save the answer; the figures update themselves. A dashboard can have pages and sections, and an authorized person can certify it as official for the organization. - **Presentation**: what you show to others. Each slide points to a chart or a finding, so if a figure changes, it also changes in the presentation. ## Alerts Middleout watches the figures that matter and emails you when they break the rule you set for them, or when they return to normal. ## Chat The main entry point: you ask in natural language and Middleout answers with data from your sources, citing where each thing comes from. Investigations, dashboards and presentations start from the chat. ## Organization, areas and roles - **Organization**: your company in Middleout. Each person acts within the organizations they are a member of. - **Organization roles**: `owner` (whoever created it), `admin` (organizes people: creates teams and decides who is in them) and `member`. - **Teams and areas**: group the people and the knowledge of each part of the company. - **Data steward**: a person authorized to curate information; for example, to certify a dashboard as official. - **Area steward**: curates the knowledge of an area. An area can have several stewards. Each person, and each assistant connected on their behalf, sees only what their permissions allow them to see. --- # Connect your sources > Which sources Middleout accepts, how each one connects and what access it needs. Middleout **only reads** your sources: it never modifies your data. Each connection asks for the minimum read access it needs. Sources are connected from the **Data** section of Middleout. ## Available sources | Source | How it connects | Personal or organization | | --- | --- | --- | | BigQuery | You grant read-only access to Middleout's service account in your Google Cloud project, or you authorize with your Google account. | Organization with the service account; personal if you authorize with your account. | | Google Sheets | You share the spreadsheet with your organization's account in Middleout, or you authorize with your Google account. | Either one. | | Google Drive | You share the folder with your organization's account in Middleout, or you authorize with your Google account. | Either one. | | GitHub | You authorize with your GitHub account and choose the repositories. Middleout studies their documentation and structure. | Personal. | | Meetings | You upload the audio recording (MP3, M4A, WAV, WebM or OGG, up to 200 MB). | Organization. | ## Personal and organization sources - **Personal**: only the person who connected it sees and queries it. Useful for working with your own data without sharing it. - **Organization**: people with access can use it, and so can the work Middleout does with nobody present (alerts, scheduled updates, investigations that refresh themselves). That work never uses personal sources. ## BigQuery There are two ways: 1. **Service account (recommended for the organization).** In the Google Cloud console for your project, you give your organization's service account in Middleout read-only roles. Middleout stores no key: access is controlled from your own IAM and you can remove it whenever you want. 2. **Your Google account.** On Google's screen you authorize the BigQuery read-only permission (`bigquery.readonly`). The source becomes personal. Each query has a limit on the data it reads, enforced by BigQuery itself, so that a question does not become expensive. ## Google Sheets and Drive - **For the organization**: share the spreadsheet or the folder with your organization's account in Middleout, as you would with a person. There is nothing else to authorize. - **Personal**: authorize the Drive read-only permission (`drive.readonly`) with your Google account. ## GitHub You authorize with your GitHub account and choose the repositories. Middleout reads their documentation (for example decisions, procedures and instruction files for agents such as `AGENTS.md` or `CLAUDE.md`) and their structure, and keeps up to date when they change. It does not run code or query the repository as a database. ## Meetings You upload the recording and Middleout listens to it and keeps what it means: the mental models, the decisions, the learnings and the open questions. ## What happens after you connect Middleout studies the source: it reads its tables and columns or its documents, uses the descriptions that already exist and writes the missing ones. You can review and edit those descriptions from each table's page. What it understands goes into the memory and is available in the chat, in artifacts and over MCP. --- # Connect Middleout over MCP > How to bring your company's knowledge to Claude, ChatGPT, Cursor or another MCP-compatible assistant. Middleout has an **MCP** (Model Context Protocol) server. Once it is connected, your assistant can query your company's memory, ask Middleout questions and record what you learn in the conversation. The MCP is optional. Middleout has its own platform, with AI included, to ask, investigate and build dashboards and presentations: you do not need a subscription to Claude, ChatGPT or any other provider. Connect the MCP if you already work with an assistant and want to bring your company's knowledge there. ## Before you start - You need a Middleout account and to be a member of an organization. - The assistant acts **as you**: it sees exactly what you see in Middleout, no more and no less. - The connection uses OAuth: you do not copy keys or passwords into the assistant. ## The server address ``` https://mcp.usemiddleout.com/mcp ``` The server uses streamable HTTP. Any client that supports remote MCP servers and OAuth can connect. ## What happens when you connect 1. You add the server address in your assistant (see the steps for each client below). 2. The assistant takes you to Middleout. If you are not signed in, you sign in with your account. 3. Middleout shows you what the assistant will be able to do in your current organization, and you choose **Allow** or **Deny**. 4. You return to the assistant, now connected. The assistant receives temporary access that renews itself. You can disconnect it whenever you want from **Settings → Connected assistants** in Middleout. ## Claude In Claude (web or desktop), add a **custom connector** from the connector settings, paste the server address and follow the authorization steps. On team or enterprise plans, an administrator may need to enable custom connectors. ## Claude Code ```bash claude mcp add --transport http middleout https://mcp.usemiddleout.com/mcp ``` Then, inside Claude Code, run `/mcp` and choose `middleout` to authorize. ## ChatGPT In ChatGPT, turn on developer mode in the apps and connectors settings, create a new connector with the server address and authorize. Availability depends on your plan, and on business accounts an administrator may need to enable it. ## Cursor Add the server to your `mcp.json` file (global in `~/.cursor/mcp.json` or per project in `.cursor/mcp.json`): ```json { "mcpServers": { "middleout": { "url": "https://mcp.usemiddleout.com/mcp" } } } ``` Cursor will ask you to authorize the first time it uses Middleout. ## Other clients Any client that supports remote MCP servers over HTTP with OAuth 2.1 (dynamic client registration and PKCE) can connect with the same address. ## Disconnect In Middleout, go to **Settings → Connected assistants** and choose **Disconnect**. The assistant loses access immediately. It also loses access if you stop being a member of the organization. ## What the assistant can do See [MCP tools](/docs/mcp-tools) for the full list. --- # MCP tools > What an assistant connected to Middleout can query and record, with the parameters of each tool. Every tool acts **as the person who connected the assistant**: it sees only what that person can see in Middleout. If something does not exist or is not visible to them, the answer is the same: not found. Memory objects are identified by a **handle**, which the search tools return and the read tools receive. ## Query | Tool | What it is for | Parameters | | --- | --- | --- | | `ask` | Ask Middleout a business question and get its grounded answer, with the sources it cites. | `question`; `investigation_id` (optional) to ask within an investigation. | | `search_knowledge` | Find what the company knows about something, by meaning or by record number. | `query`; `types` (optional); `limit` (up to 20, 10 by default). | | `read_object` | Read a whole object: its fields, its sources and, for a mental model, its causes. | `handle`. | | `get_neighbourhood` | See what surrounds an object in the graph. | `handle`; `depth` (1 or 2). | | `get_knowledge_map` | See at a glance what the company knows: its areas and the totals by type. | None. | | `get_instructions` | Read the instructions for working in a connected repository: its `AGENTS.md` and `CLAUDE.md` files. | `repository` (`owner/name`). | | `list_investigations` | List the investigations you can open, newest first. | `limit` (up to 50, 20 by default). | | `read_investigation` | Read an investigation: its question, its goal and what people have already concluded. | `investigation_id`. | The types `search_knowledge` accepts are: `metric`, `decision`, `definition`, `procedure`, `learning`, `mental_model`, `table` and `document`. `ask` is limited to **30 questions per hour per person** when used over MCP. ## Record When something the company should know comes up in the conversation, the assistant can record it in Middleout. What is recorded goes into the memory immediately, with its origin: the assistant that recorded it, on behalf of the person. | Tool | What it records | Main parameters | | --- | --- | --- | | `record_decision` | A decision: the question, what was chosen, the alternatives and the assumptions. If another decision on the same question is recorded, it replaces the previous one, which is kept as history. | `question`, `outcome`, `alternatives`, `assumptions`, `rests_on`, `scope`. | | `record_teaching` | A teaching: a definition, a rule or a formatting preference. | `statement`, `subject`, `confirmed`, `scope`, `kind` (`definition`, `rule` or `formatting`). | | `record_relation` | A relation between two objects in the memory, with the quote that backs it. | `source`, `target`, `relation`, `quote`, `confirmed`. | | `record_question` | An open question nobody has answered yet. | `statement`, `scope`. | | `record_evidence` | Evidence for or against something the memory already contains. | `statement`, `stance` (`supports` or `contradicts`), `bears_on`, `scope`. | `scope` says who the record is for: `org` (the whole organization), `team` (your team) or `user` (only you). ## Good practices for agents - Before answering about the business, search with `search_knowledge` or ask with `ask`: the company may have an agreed definition that differs from the usual one. - Cite the handles you used, so the person can open them in Middleout. - Record only what the person confirmed or decided in the conversation, and set `confirmed` when they did so explicitly. - If a tool says something does not exist, the person may not have access to it. Do not assume it does not exist for the whole company. --- # Security and data handling > How Middleout accesses your sources, where it runs, which models it uses and how it keeps each organization separate. ## Read-only Middleout never modifies your data. - **BigQuery**: read-only permission (`bigquery.readonly`), or read roles you give your organization's service account from your own IAM. - **Google Sheets and Drive**: read-only permission (`drive.readonly`), or shared read access with your organization's account. - **SQL databases**: each query runs in a read-only transaction. ## Limits per query So that a question does not read more than it should: - Each query has a cap on the data it reads, enforced by BigQuery itself. - A query returns up to 10,000 rows; a dashboard chart, up to 1,000; an export, up to 50,000. ## Separation between organizations Every piece of data belongs to an organization. A query without a defined organization fails instead of returning rows, and the work Middleout does with nobody present uses only the organization's sources, never personal ones. Within an organization, each person sees only what their permissions allow. An assistant connected over MCP sees exactly the same as the person who connected it. ## Credentials - With the BigQuery service account, Middleout stores no key: access is granted and removed from your IAM. - Authorizations with Google or GitHub can be revoked at any time from your account in those services. - Assistants connected over MCP use OAuth with temporary access that renews itself. They never receive your password, and you disconnect them from **Settings → Connected assistants**. ## Where it runs Middleout runs on Google Cloud (Cloud Run and Cloud SQL), in the `us-east1` region (United States). ## AI models Middleout uses Gemini models through Google Cloud's Vertex AI, with Vertex AI's global location. The models are used to converse, to understand your sources and meetings, and to search by meaning. ## Personal data The processing of personal data on the Middleout site is described in the [privacy policy](/privacy).