The knowledge base MCP server that keeps AI agents on your procedures
Neuron MCP connects your SOPs, policies and how-to knowledge to Claude, ChatGPT, Cursor, VS Code or Claude Code. Your assistant answers from the current article and links to it. Your agents draft what is missing and log the gaps. A person approves every publish.
One URL per workspace. Sign in. Done.
What is Neuron MCP?
Neuron MCP is the Model Context Protocol (MCP) server built into every Neuron workspace, our knowledge base for SOPs and training. MCP is the open standard AI assistants use to call tools in other software. With Neuron MCP connected, your assistant can search and read your company's own procedures directly, instead of you pasting documents into a chat and hoping they are the current version.
It is MCP-native, not a bolt-on. Articles are stored as structured Markdown with an owner, a version and a review date, so what reaches your AI is the same article your team reads, with the context that says how far to trust it. You choose the model. Neuron supplies the knowledge, the search and the guard rails.
10
tools
3
access levels per person
0
tokens to copy
0
articles published by AI
How it works
From a question to a published procedure
A model on its own answers from general knowledge. With Neuron MCP it answers from your articles, and what it can't answer becomes work for an editor instead of a guess.
- Step 1
Ask
A person, or an agent working on their behalf, asks a question. The assistant searches Neuron with a few distinctive keywords and your category, tag and type filters.
- Step 2
Read
It opens the whole article as Markdown, with the owner, the version and the review date, so it knows how current the answer is.
- Step 3
Cite
It answers with a link to the article. Wording you have locked, such as legal or compliance text, is quoted exactly instead of paraphrased.
- Step 4
Capture
When nothing covers the question, it says so and can log the gap. With drafts access it can write the missing article as a draft in your format.
- Step 5
Approve
An editor reviews the draft in Neuron and publishes it. Readers are notified, and acknowledgments are collected if the article needs them.
Example
Ask a question. Get the procedure, not a guess.
An illustrative conversation. The assistant calls Neuron MCP tools for the content, then answers from what it read.
An employee wants to expense a $750 software subscription. What do we do? Quote the policy.
This is covered by Expense reimbursement procedure (SOP, version 4, reviewed last month). The wording is locked, so here it is exactly:
“Claim expenses within 30 days of purchase. Anything over $500 needs written approval from the department head before it is reimbursed.”
At $750 this is over the limit, so the employee needs their department head's written approval before submitting the claim. Source: /a/expense-reimbursement
I found nothing about reimbursing travel booked on a personal card. Want me to log that as a gap?
Yes, and draft it from the notes I pasted above.
Logged the gap and saved Reimbursing travel on a personal card as a draft, tagged with your existing tags. It is waiting for an editor to review. Nothing has been published.
More things to ask
- “What's our expense reimbursement procedure? Quote it.”
- “Which articles cover data security?”
- “Summarise the onboarding SOP for a new hire.”
- “Draft a how-to from these notes and tag it properly.”
- “Is there anything on requesting time off? If not, log a gap.”
- “Which categories do we have for HR content?”
Agentic by design
Give your agents the rule book, and a safe place to write
Agents do better work when they can check how your company does things, and they are safer when the only thing they can write is a draft. Neuron MCP gives them both.
Answer from the SOP
“What's our process for approving a purchase? Quote the exact wording and link the article.”
The assistant searches, opens the procedure and answers from it with a citation, instead of guessing from general knowledge.
Turn notes into a draft
“Turn these call notes into a how-to in our format and save it as a draft.”
It checks your categories and canonical tags, writes the article with the right fields, and saves it as a draft for an editor.
Find what's missing
“Do we have a procedure for requesting time off? If not, log it as a gap.”
It searches, tries a synonym before giving up, and records the missing topic with its evidence in your gap backlog.
Keep a draft moving
“Update my onboarding draft with the three new steps from this thread.”
It replaces the draft's content as a new draft version. Published articles are out of reach: those are edited in Neuron.
Ground an agent in your rules
“Before you reply to this ticket, look up our refund policy and follow it.”
Coding agents, support agents and research agents check the real procedure first, so their work follows how your company does things.
Told how to behave
The server tells every connected app what the knowledge base is, to cite the article it used, to quote locked wording exactly, to use your canonical tags, and that anything it writes is a draft.
Tools
Everything your assistant can call
Six tools read and four need drafts access. A connection is only shown the tools it may use, so a read-only assistant never sees the draft tools.
| Tool | Access | What it does |
|---|---|---|
| search_kb | Read | Keyword search over published articles, filtered by category, tag or type. Returns up to 10 results with title, summary, a matching snippet and a link. |
| get_article | Read | One whole article as Markdown with its fields: type, category, tags, owner, version and review date. |
| search | Read | The same search in the shape ChatGPT expects, for deep research and company knowledge. |
| fetch | Read | The article behind a search result, in the shape ChatGPT expects. |
| list_categories | Read | Every category with its path, name and description. |
| list_tags | Read | Your canonical tags with their descriptions and aliases. |
| create_draft | Drafts | Saves a new article as a draft, checked by the same rules as an import and labelled with the app that wrote it. Never published. |
| update_draft | Drafts | Replaces the content of an article that is not published yet, as a new draft version. Published articles are refused. |
| log_gap | Drafts | Records a topic the knowledge base doesn't cover, with the reason and the evidence, in the gap backlog. |
| propose_tags | Drafts | Suggests canonical tags for an article without saving anything. Relies on Neuron's own AI features, which are switched off at launch. |
There is no tool that publishes, edits a published article or deletes anything. Read-only tools are marked as such, so apps like Claude can run them without asking each time and ask before the others.
Access levels
An admin decides who connects an AI, and how far it goes
MCP access is set per person, on top of their role. A connection can do the lowest of what the app was granted and what the person is allowed.
Off
The default for new members
No AI app can connect as this person. Sign-in from an AI app is refused.
Read
Readers, editors and admins
Search and read, and nothing more. The assistant sees what the person sees: a reader's connection gets published articles only.
Drafts
Editors and admins
Read, plus creating and updating drafts and logging gaps. The person also has to tick a box when they connect each app.
Why MCP-native
Shared drives and copy-paste vs Neuron MCP
| Topic | Shared drives and copy-paste | Neuron MCP |
|---|---|---|
| Where the answer comes from | Whatever someone pasted into the chat, or the model's general knowledge. | The current published article, with a link back to it. |
| Exact wording | Paraphrased, even where the precise words matter. | Locked articles are quoted exactly. |
| What the AI can change | Anything in a shared folder it was given. | Drafts only. A person publishes, and nothing is deleted. |
| Who sees what | Whoever has the file or the link. | What that person's role sees, checked on every request. |
| When knowledge is missing | The model fills the hole with a confident guess. | It says the topic isn't covered and logs a gap for an editor. |
| Which model you use | The vendor's built-in assistant, if there is one. | Your own: Claude, ChatGPT, Cursor, VS Code or Claude Code. |
Connect your AI
Works with the assistant you already use
Each person's Connected apps page shows their workspace's URL and the steps for each app. Paste the URL, sign in to Neuron and approve. Neuron supports standard MCP sign-in discovery and PKCE, so compliant clients need nothing else.
Claude (web, desktop and mobile)
Add a custom connector with your workspace's URL, sign in with your email and approve.
ChatGPT
Add your workspace as a connector. Neuron's search and fetch tools fit deep research and company knowledge.
Claude Code
One command with your workspace's URL, then sign in from inside Claude Code.
Cursor
Add the URL to mcp.json and complete the browser sign-in.
VS Code
Add an http server to mcp.json; VS Code opens the sign-in for you.
Scripts and headless agents
Use an API token an admin creates, sent as an Authorization: Bearer header. It follows the same access rules.
Articles are also available as raw Markdown and as a JSON Lines export over plain HTTPS, for pipelines that don't speak MCP.
Security and permissions
Your AI gets the handbook, not the keys to it
Sign in, no token to copy
Connect by URL with standard MCP OAuth sign-in (PKCE). You sign in to Neuron with your email and approve the app. There is nothing to request, copy or paste.
Off until an admin allows it
Every person has an MCP access level: off, read or drafts. New members start at off, and nobody can raise their own level.
Sees what the person sees
A connection never has more access than the person behind it. A reader's assistant gets published articles only.
Drafts never publish themselves
Writing needs an editor, an admin's permission and a tick on the consent page. Every write is a draft, labelled with the app and recorded in the audit log.
Cut off on the next call
Disable a person or switch their MCP access off and every connection stops at its next request. Unused connections lapse after 30 days by default.
Approved apps only
Only AI apps on your workspace's approved list can connect. An admin edits the list, and anything else is refused before sign-in.
Each workspace has its own database, so a connection to one workspace can't reach another customer's content. The server itself calls no AI model: the AI runs in your own Claude or ChatGPT account.
Included with every Neuron plan
Neuron MCP comes with Starter (from $39 a month, 10 users included), Team and Business: searching, reading and drafting. No per-query fees and no AI markup, because you use the assistant you already pay for.
Compare Neuron plansFAQ
Neuron MCP questions
What is a knowledge base MCP server?
MCP (Model Context Protocol) is an open standard that lets AI assistants call tools on other services. A knowledge base MCP server like Neuron MCP lets your assistant search and read your company's own procedures and policies directly, and cite them, instead of you pasting documents into a chat.
Which AI assistants work with Neuron MCP?
Claude (web, desktop and mobile), ChatGPT, Claude Code, Cursor and VS Code each have setup steps in Neuron. Which apps can add custom connectors, and on which of their plans, is up to each app.
Does Neuron run its own AI model?
No. Neuron doesn't bundle a model at launch. The AI that talks to Neuron runs in your own Claude or ChatGPT account, and Neuron supplies the knowledge, the search and the tools it works from.
Do I need an API key?
No. You paste your workspace's MCP URL into the app, sign in to Neuron with your email and approve. API tokens are still available for scripts and agents that can't do a browser sign-in.
Can an AI agent publish, edit or delete published articles?
No. With drafts access an agent can create a draft, update an article that is not published yet, and log a gap. It can't publish, can't change a published article and can't delete anything. A person reviews each draft and publishes it in Neuron.
Is the search semantic?
No, it is keyword search. Neuron tells the assistant so: use one to three distinctive words, and try a synonym before concluding a topic isn't covered. When no article has every word, results matching some of them are returned and marked as such.
Is our content sent to the AI company?
Articles your assistant opens are sent to the provider of that assistant, as with anything you put in a chat. Neuron shows each person how to switch off model training in Claude and ChatGPT before they connect, but it can't enforce that setting.
Which Neuron plans include MCP?
Every plan includes searching, reading and drafting over MCP. Neuron is coming soon; join the early access list to hear when it opens.
Put your procedures where your AI can find them
Neuron MCP is coming soon. Leave your email and we'll tell you the day it opens.