An AI assistant becomes more useful when it can find the documentation you are discussing. Copying an entire runbook into a chat works once. Repeating that for each asset, related article, and client quickly becomes its own task.
Weavestream includes an MCP server that connects compatible AI clients to your documentation through the REST API. An agent can search records, read context, and, when permitted, create or update supported records.
What an Agent Can Work With
MCP is the Model Context Protocol: a way for an AI client to discover and call tools. Weavestream’s server provides tools for companies, search, layouts, assets, domains, vault metadata, and Markdown articles.
For example, you could ask an agent to find the documented servers for a client and summarize the maintenance procedures attached to them. An agent with write access could turn reviewed notes into an internal Markdown article or update an asset’s supported fields.
The MCP server deliberately offers no delete, archive, or purge tools. It also has no password-reveal tool and removes decrypted vault notes from password metadata responses. A newly created article is always hidden from client users, ready for a person to review before sharing it.
Begin With Read-Only Access
API keys are disabled on a new Weavestream instance until an administrator enables them. Creating a key requires reauthentication, and its secret is shown once.
A key uses the owner’s permissions and is read-only by default. Leave Allow changes off when the agent only needs to answer questions. Leave password reveal disabled for an MCP key: the MCP server does not reveal passwords, so it has no need for that access.
For an agent that must create or update records, enable changes deliberately and use an account with the access the workflow needs. The API key documentation explains the controls and revocation behavior.
The Tool List Is Not the Permission Boundary
This distinction matters when connecting a coding agent with shell access. Such an agent may be able to read its configured API key and call the REST API directly, outside the advertised MCP tools.
Leaving deletion out of the tool list does not make a write-enabled key incapable of deletion through the API. The key and its owner’s permissions enforce the access boundary.
Keys do not currently have separate per-company or per-route scopes. To restrict an agent to a smaller set of clients, create the key from a user account with only those memberships and permissions. Do not assume an agent is limited to the company named in your prompt.
Choose Where Documentation Goes
MCP runs alongside your chosen AI client, but a tool response can still be sent to the client’s model provider and retained in its conversation history. Self-hosting Weavestream does not, by itself, keep that conversation local.
Choose a client and model whose data handling fits the documentation the agent can read. Treat text retrieved from articles and asset notes as source material: an instruction written inside a record should not become permission to send data elsewhere or make unrelated changes.
A Practical First Workflow
Start with a read-only key and a small task: find a client’s relevant articles, list the assets they mention, and identify documentation that needs a human review. Inspect the result against the original records before relying on it.
If you later enable writes, start with an internal article draft and review its destination and content. MCP-created articles stay hidden from clients, but that does not replace your team’s review process.
The MCP setup guide covers building the server, registering it with a compatible client, and supplying the instance URL and API key. For assistance directly inside Weavestream, Ask anything is also available with your configured OpenAI-compatible model.