Skip to main content
The Within MCP connects your own AI client to your Work Brain (your Process Index), so you can analyze your processes and build skills and agents in the tools your team already uses. This page covers when to reach for it instead of Advisor, and how to work with it for each job. Three parts to keep straight:
  • Advisor creates inside Within. It answers questions across your Work Brain and generates reports, diagrams and skills from your process data.
  • The Within MCP retrieves from Within. It is a read-only connection that gives an AI client access to the process data you already have permission to see.
  • Your AI client acts outside Within. Once connected, it combines Within data with the other connectors it already has, and creates files in your own tools.
MCP stands for Model Context Protocol, an open standard for connecting an AI client to a data source. Within publishes a server. Your client connects to it.

​
Advisor or the MCP?

The rule: use Advisor when the analysis, and what it produces, should stay in Within. Use the MCP when the work needs files or other systems your AI client is connected to, or has to happen inside your client.

Use Advisor to

  • Find your biggest opportunities across the Work Brain. It reasons across every process in scope and keeps the result in Within.
  • Produce a report, diagram or analysis to share from Within.
  • Find out which skills are worth building. It ranks candidates by how often the work happens, the active work time, its complexity and the total time it takes over a year.

Use the MCP to

  • Check a process against another system. Your client already has connectors for your ticketing, project and file tools.
  • Produce a file that draws on your other tools. Your client creates it, combining Within data with its other connectors.
  • Build an agent, in a coding agent. The Within ADK is served through the MCP.
To build a skill for one bounded task, use either: Advisor when you want it to recommend and package the skill, or your client when the skill needs files or connectors it already has. See Build a skill. To change a process or reorganize your Work Brain, use AI Chat in Within. The MCP is read-only. For the detail behind each, see Strategic analysis with Advisor, Advisor: analyze processes and Build an agent.

​
How they compare

​
Analyze with the MCP

Your client can search, read, explore and analyze your Within data, but not change it. Analysis belongs in the MCP when it needs something Advisor can’t reach:
  • Data in another system. Teams often ask Within to build native integrations with the ticketing and project tools they already run. The MCP is the working path today: check a documented process against the tickets, projects or files your client is connected to, and find where the work deviates from it.
  • A current-state read before you build. Before you design a skill or an agent, read the observed work: pain points, bottlenecks, the systems in use, and what already works well.
  • Output that has to live in your tools. A control matrix, a process export or an executive deck, built where your team will use it.
For analysis across your whole Work Brain — where your biggest opportunities are, and where to deploy agents first — start in Advisor. Prompts for each kind of analysis are in Example prompts.

​
Build a skill

A skill is a package of files — instructions, context and guardrails — that teaches an AI client one bounded task, built from how your team actually does that work. You install it in the AI platform your team uses. You can build one in Advisor or in your client through the MCP. Both end in a package you install in your AI platform. They differ in where you build it and what it can draw on.

​
In Advisor

Use Advisor when you want it to find the opportunity as well as build the skill, and when the analysis should stay in Within.
1

Ask for recommendations first

Ask Advisor where a skill would save the most time, across a value stream or a set of processes. It ranks candidates by how often the work happens, the active work time, its complexity and the total time it takes over a year.
2

Approve the scope

Advisor proposes a name and scope for the skill and asks before it builds anything. Change the scope in plain language with Suggest another way, or approve it with Proceed.
3

Review the package

Preview each file in the conversation. Advisor also generates a diagram of the workflow before and after the skill.
4

Download and set it up

Download the package as a ZIP and set it up in your AI platform, for example as a project in Claude.
If a process isn’t a good candidate — too much judgment to codify, or too little detail in your Work Brain — Advisor says so rather than generating a weak skill.
Skill generation in Advisor is in beta. A skill you generate is private to you: to share it, send the package, and set it up in your platform yourself.

​
In your AI client

Use your client when the skill needs files or connectors your client already has, or when your team builds and keeps its skills there. Ask your client to build a skill for a named process. It follows the build method Within serves through the MCP: it gathers the evidence for the process, asks you to approve the scope, then hands you the package with the results of a quick test before you install it.
Decide where the skill stops before you decide what it produces. A skill that knows when to hand back to a person is easier to review and easier to trust.

​
If your team builds skills regularly

  • Give every skill an owner, a version and an evaluation set when you distribute it.
  • Have a small review group decide which skills are safe to share and who maintains each one.
  • Keep approved skills in one place, such as a shared repository, and retire the ones that go out of date.
The full approach for business users is in the non-technical track of the developer docs.

​
Build an agent

When one bounded task isn’t enough — the inputs vary too much for a fixed path, or the work spans several tasks and tools — you’re building an agent. Build it in the coding agent your team already uses, connected to the MCP. The Within ADK, the build method, is served through the MCP, so every build uses the current version. Choose the smallest shape that does the job: a skill, an Integration Flow, a managed agent or a context bundle. See What you can build. If you build, release and maintain agents for your team — an AI or automation lead, or the systems analyst who owns your team’s automations — start every build from the captured process, not from a blank prompt.
1

Pull the process before you design anything

Ask for the steps, owners, systems, decisions and exceptions. That’s where the agent’s boundaries come from.
2

Turn it into a spec

Specify, build and validate with the Within ADK. See Build an agent with the Within ADK.
3

Test against the exceptions, not the happy path

The variations Companion observed are your evaluation set. See Validate an agent before release.
4

Come back when the work changes

Companion keeps capturing after launch. Compare the latest observations with the evidence baseline you built on. See Maintain and improve agents.
The MCP is read-only, and your client sees only what you can see in Within. If the process is wrong, fix it in Within and pull it again. Don’t patch it in the agent.

​
What to know before you connect

  • The MCP is read-only. Your client pulls what it asks for: your process hierarchy, process details, and the timelines and observations from Companion sessions you already have access to. Changes to a process or your Work Brain happen in Within. See limitations.
  • Your permissions do not change. MCP requests run through the same checks as Within in your browser, whichever way you sign in. You cannot reach a workspace you are not a member of. See Security and architecture.
  • Your AI client supplies the model. The MCP does not connect to an external model itself. Your client does, using your own subscription and your organization’s agreement with that provider.
  • What Within keeps. An audit log of the tools your client called and the arguments it passed, plus any feedback you submit through the feedback tool. Within does not see your conversation with your AI client.
  • Some capabilities belong to your client. Creating files, citing sources, and running work on a schedule are things your AI client does. Support varies by client and plan.
  • One connection covers one workspace. If you belong to more than one, switch by reconnecting, or by using a key generated in the other workspace. See Switching workspaces.

​
Connect

Setup takes a few minutes and is done once. Full instructions for each client are in the developer docs.
1

Add the Within connector in your AI client

Sign in to Within when the client asks. Steps differ by client. See the install guide for yours.
2

Ask for a process by name

Your client pulls the steps, owners, systems, and evidence you have access to.
3

Analyze or build from it

Cross-check it against another connector, export it to a file, or turn it into a skill or an agent.
Supported clients are Claude, Claude Code, ChatGPT, Codex, Cursor, Gemini CLI, GitHub Copilot and Microsoft Copilot. For Claude and ChatGPT, install from their official connector directories. For authentication options, see Authentication.

​
Example prompts

Replace the process names with your own. Creating files depends on your client.

​
Analyze

“For each step in our ‘Access Review’ process from Within, match it against the tickets in my ticketing export and flag steps with no supporting evidence.”
“Using my Within and [project tool] connectors together, compare our documented ‘Bug Triage’ process with the actual tickets from last month and list where the team is deviating from the process.”
“List every system the [process] touches in Within, step by step, with what is read or written. Mark which ones this project already integrates with and which would be new.”

​
Create files in your tools

“Pull the full detail for our ‘Month-End Close’ process from Within and write every step, owner, system, and evidence into an Excel file, one row per step.”
“Pull the ‘Procure-to-Pay’ process and its pain points from Within and build an executive slide deck with a process diagram and a quick-wins slide.”
“For our quarterly access-review process in Within, pull each control step and its evidence and build a control matrix spreadsheet: control → owner → evidence.”

​
Build skills and agents

“Fetch the ‘Invoice Discrepancy Resolution’ process from Within and set up a reusable skill that automates the three most repetitive manual steps.”
“Connect to Within and answer my team’s questions about how any of our documented processes work, always citing the source.”
“Pull the ‘Invoice Discrepancy Resolution’ process from Within with every step, system, decision and exception. Draft an agent spec: the trigger, the inputs and where each comes from, the steps the agent owns, where it stops for a person, and the outputs.”
“From the Companion observations of [process] in Within, list every variation and exception, including missing, late or changed inputs, and write them as test cases with the expected outcome.”
“Compare the [process] in Within with the agent in this project. Which steps, systems or exceptions in the process does the agent not handle?”

​
Where to go next

Build an agent with the Within ADK

Specify, build and validate an agent in your coding agent.

Strategic analysis with Advisor

Find your biggest opportunities across the Work Brain.

Advisor: analyze processes

How to run an Insights analysis, step by step.

Within MCP developer docs

Setup, concepts, security, guides, and tool reference.

Using AI Chat

Edit process details and reorganize your Work Brain.

Advisor prompt library

Ready-to-use prompts by persona and use case.
Need help? Use the in-app chat — click the chat bubble in the bottom-right corner of Within (staffed 24×5) — or email support@within.ai.