watchdep
Sign inSource
Docs / Connect / 15

Agents (MCP)

Section
Connect
Page
15 of 16
Reading time
4 min

watchdep is an MCP server. Connect Claude Code, Claude Desktop or any MCP client that supports remote servers with OAuth, and it can list findings, open one, look through a part’s inventory, find which applications use a package and explain a level. If you allow it, it can also record a decision. An agent always acts as the person who connected it, with that person’s role, in one workspace, and everything it does is in the log as an agent’s action.

Connect Claude Code

The MCP server is your watchdep server’s address followed by /mcp. In Claude Code:

claude mcp add --transport http watchdep https://watchdep.example.com/mcp

Settings, Agents shows your watchdep server’s address and this command, ready to copy.

Connect another MCP client

Give the client https://watchdep.example.com/mcp (your watchdep server’s address with /mcp) as a remote MCP server. The client registers itself with watchdep and then opens watchdep in your browser to sign in. Any client that supports remote servers with OAuth works the same way.

Signing in and choosing the scopes

  1. The client opens your copy in the browser. You sign in as usual, with two-step sign-in if you set it up.
  2. The consent screen shows the client’s name (which the client chose for itself, so read it as a claim), the address it will send you back to (marked in deep orange when that is not your own computer), you, your role and the current workspace. If anything about the request is wrong, you get a page explaining it; watchdep never sends you anywhere without a click.
  3. Read is always given. Decide is given only if you tick it; it is off by default. If your role cannot decide (a viewer), the screen says ticking it would change nothing.
  4. Choose Connect, or Deny.

The connection is to the workspace that was current on the consent screen. To connect an agent to another workspace, switch to it first, then connect again.

The six tools

Tool Scope What it answers
list_findings read Open findings, loudest first. Filters by application, level, lane, state and text in the id or summary; up to 100.
get_finding read One finding with every place and dependency path, the code check’s evidence, the fix, pull requests and the full log, the same as the API’s finding.
get_part_inventory read A part’s production inventory (or its last one): files read and not read, runtimes, and packages with versions, scope, relation and a dependency path. Says “not looked yet” when the part has sent none.
find_package read Every application and part whose production inventory has a package of that name, with versions.
explain_level read The finding’s level and its inputs (KEV date, EPSS and the workspace’s threshold, CVSS, the source’s own label, scope, malicious, lane) with the ranking rules in words.
record_decision decide Acknowledge, mark not affected (with a reason), or accept the risk until a date (with a reason, at most a year away), as you.
  • An agent connected without decide is not offered record_decision at all.
  • record_decision follows your role: a viewer’s agent is refused even with decide ticked.
  • A decision through an agent follows every rule a decision on the screen does. See Decisions and Levels and lanes.
  • Everything an agent does is in the log as the person, through the agent, for example “D. Kim, through Claude Code”.

What an agent can never do

No tool does any of these, whatever scopes were ticked:

  • read any workspace but the one it was connected to;
  • open pull requests (fix pull requests are opened only by watchdep fix in your own CI, and only a person merges);
  • change alert settings or fix settings;
  • reach workspace settings, members, tokens or connections.

Seeing and disconnecting agents

Settings, Agents lists each live connection: the agent, the person, the scopes, when it was connected and when it was last used. You see and disconnect your own; owners and admins see and disconnect everyone’s. Disconnecting is immediate: the agent’s tokens stop working at once.

A person who leaves the workspace takes their agents with them.

Limits

What Limit
Client registration 20 in 10 minutes from one address
The token endpoint 60 requests a minute
An authorization code Lasts ten minutes and works once, right or wrong
An access token Lasts an hour
A refresh token Lasts 30 days and is used once: every refresh gives a new pair, and using one a second time ends the connection