Skip to content

MCP for advisors ​

An advisor can connect an AI assistant, such as Claude Code, to their DutyGraph workspace. The assistant can then add accounts, import CSV files, start research, read the Company Snapshot and the Company Repo, and prepare outreach.

The advisor MCP server is a second way to do what the app does. It is not a second set of rules. Each tool calls the app's own route, as the advisor who owns the token.

Create a token ​

Step 1

Open the panel ​

Sign in as an advisor. Open Settings › MCP access, then Connect an AI assistant (MCP).

Step 2

Name it and choose scopes ​

Enter a name, for example "Claude Code on my laptop". Choose the scopes (below).

Step 3

Choose when it expires ​

Choose 30 days, 90 days (the default), 1 year or No expiry.

Step 4

Copy it once ​

Select Create token. Copy the token now. It starts with dgp_. DutyGraph shows it once and keeps only its SHA-256 hash.

The token list shows each token's name, scopes, creation date, last use and expiry. Revoke stops a token at once. A token cannot create or revoke tokens.

The shared sample login cannot create tokens, because every visitor shares that workspace.

Scopes ​

ScopeWhat the assistant can do
Read (always on)See companies, accounts, research status, the Company Snapshot, Company Repo files and outreach drafts
WriteAdd accounts and contacts, import CSV files, start Preview or Snapshot research, prepare and edit outreach drafts. Write does not send email.
Send email (off by default)Send a prepared outreach draft. Turn it on only if you review each draft before you tell the assistant to send it.

A fourth scope, request_agents, belongs to employees. It cannot be combined with these. See Request from your AI.

Tools ​

ToolScopeWhat it does
list_companiesreadCompanies in the workspace
list_accountsreadAccounts with stage, research level and primary contact
get_accountreadOne account: stage, engagement, contacts, research authorizations, drafts, recent events
get_research_statusreadThe latest research runs and whether a research source is ready
get_company_snapshotreadThe findings the Company Snapshot shows, with their sources
get_company_repo_indexreadThe Company Repo index: each file's path, title and state
read_company_repo_filereadOne Company Repo file, with its state and content
list_outreach_draftsreadDrafts, recipient, sender gaps, quota and research signals
add_accountwriteOne prospect, with the same duplicate checks as Accounts › Add account
import_accounts_csvwriteA CSV of accounts. Previews unless apply: true.
add_contactwriteA contact on an account
run_researchwriteStarts Preview or Snapshot research and returns the run id
import_roster_csvwriteA company's team list. Previews unless apply: true. Invites nobody.
draft_outreachwritePrepares the outreach note from the app's template and research signals
edit_outreach_draftwriteChanges the subject and body. The content checks run again.
send_outreachsend_emailSends a saved draft through the app's Send path

read_company_repo_file can read a saved agent manifest's Repo file (manifests/<slug>.md), so an assistant can read the manifest the builder made.

Connect ​

The endpoint is <your DutyGraph origin>/api/mcp. Send the token as Authorization: Bearer <token>.

Claude Code

Run the command the panel shows:

bash
claude mcp add --transport http dutygraph https://<your-dutygraph-host>/api/mcp \
  --header "Authorization: Bearer <token>"

Claude Desktop

Claude Desktop needs the local mcp-remote bridge and Node.js. Add a server to Claude Desktop's configuration in Settings › Developer, then restart Claude Desktop:

json
{
  "mcpServers": {
    "dutygraph": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://<your-dutygraph-host>/api/mcp",
               "--header", "Authorization:${DUTYGRAPH_AUTH}"],
      "env": { "DUTYGRAPH_AUTH": "Bearer <token>" }
    }
  }
}

DutyGraph shows this configuration in the employee's Connect your AI panel. The advisor panel shows only the Claude Code command. The endpoint and header are the same.

Other assistants

Add a Streamable HTTP MCP server with the URL <origin>/api/mcp and the header Authorization: Bearer <token>. The server is stateless: it answers POST with JSON.

Claude web custom connectors and ChatGPT need OAuth sign-in. That is Planned.

Safety rules ​

  • Same user, same rules. A token acts as its advisor, inside the advisor's workspace. Another workspace's company returns "not found".
  • Audited, marked as MCP. Audit events and account events written during a tool call name the token's user as the actor, with via: "mcp" and the token id. Creating and revoking a token are audit events too. Read-only tool calls leave no audit record yet.
  • No contact with employees. No tool invites or emails a company's people.
  • Outreach goes through the app's checks. send_outreach uses the same Send path as the app: sender identity, suppression lists, the footer and unsubscribe, approval and limits. If a check warns, the tool refuses and says "Needs approval in the app".
  • No secrets. No tool returns provider keys, email settings or private links.
  • No decisions. No tool can approve, change or deny an agent request.

A missing, revoked or expired token gets HTTP 401. Failed token checks count against a per-address limit.

The other MCP surfaces ​

INFO

Status. Built: the advisor MCP server, tokens with read, write and send email scopes, the 16 tools, audit marking and token revoke. Simulated: none on this page. Planned: OAuth sign-in for web and mobile connectors, and an audit record for every read-only tool call.

Source: gitbook/product/connect/mcp.md in the DutyGraph repository.

DutyGraph. Pages label each feature Built, In review or Planned.