DokuTrak

dokutrak-mcp

The open MCP connector for DokuTrak: let your agent chase the documents.

DokuTrak collects documents from your clients on your behalf: you send a request, the client uploads through a secure link, the files are reviewed, and silent clients get reminded. This connector puts that loop inside the agent you already work in, so "where does the Dupont file stand?" is answered without leaving Claude.

The connector is a thin, stateless client of the DokuTrak API. It holds the Agent Connection you give it, stores nothing on disk, keeps no cache, and duplicates no rule: what your agent may and may not do is decided by the service, and refusals come back as tool errors with the service's own explanation.

Install

You need a DokuTrak workspace and an Agent Connection, issued from Settings → Connect an agent in the DokuTrak app. That screen hands you a paste-ready configuration with your key already in place; the instructions below are the same thing, by hand.

The key is read from the environment variable DOKUTRAK_API_KEY. It is never taken from the command line.

Install in Claude Desktop (one click)

1. Download dokutrak.mcpb, the connector packaged as an MCP Bundle.
2. Double-click it, or drag it onto Claude Desktop (Settings → Extensions).
3. Click Install and paste your Agent Connection key when asked. Claude Desktop masks it, stores it securely and passes it to the connector as DOKUTRAK_API_KEY.

The bundle carries its own copy of the connector and runs on the Node.js built into Claude Desktop: no npx, no configuration file to edit. npm run bundle builds the same file from a clone.

Claude Desktop, by hand

Open the configuration file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the server under mcpServers (create the object if the file is empty):

{  "mcpServers": {    "dokutrak": {      "command": "npx",      "args": ["-y", "dokutrak-mcp"],      "env": { "DOKUTRAK_API_KEY": "dk_live_…" }    }  }}

Restart Claude Desktop. The DokuTrak tools appear in the tools menu of a new conversation.

Claude Code

claude mcp add dokutrak -e DOKUTRAK_API_KEY=dk_live_… -- npx -y dokutrak-mcp

Then /mcp inside Claude Code lists dokutrak and its tools.

Claude Code plugin: tax-season document chase

/plugin install dokutrak --marketplace Crackx17/dokutrak-mcp

The plugin adds the hosted connector (you sign in to DokuTrak on first use) and the tax-season-document-chase skill: give Claude your client list, it builds each client's checklist by return type (W-2, 1099, Schedule C, rental, K-1, retiree…), shows you the whole batch to confirm, prepares one request per client for you to confirm in DokuTrak, sends them, and later reports who is still missing what. Nothing goes out without your confirmation.

claude.ai and ChatGPT (hosted)

Add https://app.dokutrak.com/api/mcp as a custom connector. You sign in to DokuTrak (or create an account) during the OAuth step. The hosted connector prepares a request, sends it only after you confirm it in DokuTrak, and tracks progress. It never returns file content. Downloading files stays with this local connector. Setup: dokutrak.com/connectors.

From a clone, before the npm release

git clone https://github.com/Crackx17/dokutrak-mcp.gitcd dokutrak-mcpnpm ci && npm run build

Then point the client at the built file instead of npx:

{  "mcpServers": {    "dokutrak": {      "command": "node",      "args": ["/path/to/dokutrak-mcp/dist/cli.js"],      "env": { "DOKUTRAK_API_KEY": "dk_live_…" }    }  }}

or, for Claude Code: claude mcp add dokutrak -e DOKUTRAK_API_KEY=dk_live_… -- node /path/to/dokutrak-mcp/dist/cli.js.

The skill

skills/dokutrak/SKILL.md teaches the agent the three everyday uses — ask a Client for documents, know where a request stands, chase on rejected files — and how to connect. It is what a DokuTrak user installs alongside the connector:

npx skills add Crackx17/dokutrak-mcp        # the open agent-skills installer# or by hand, for Claude Code / Claude Desktop:cp -r skills/dokutrak ~/.claude/skills/dokutrak

Configuration

Tools

Four tools, one round trip: ask, chase, know, collect.

create_request

Creates a Document Request and sends it, in one call, so nothing is left created but unsent. Takes the client's email, a deadline (YYYY-MM-DD or an ISO datetime), the checklist of documents wanted, and an optional title and message. The email goes to the recipient given here and to nobody else; the client uploads through the secure link it contains. Under the hood this is the same two-step the DokuTrak app performs: create with sendEmail: false, then send. If the send fails, the error names the created request, which stays visible in the dashboard.

Nothing goes out without your yes. The email to a real client cannot be recalled, so the tool tells the agent to show you the recipient, the deadline, the checklist and the message, and to wait for your confirmation. The tool is also flagged so that the client asks you before every call: Claude Code prompts each time, even in auto or bypass mode, and Claude Desktop treats it as a tool that always needs approval. A client that ignores these flags is left with the instruction to the agent alone.

request_replacement

Chases the client on the rejected files of a request: flags them, moves the request back to awaiting the client, and returns it to the automatic reminder cadence. This call sends no email by itself; the reminders do, and DokuTrak has no way to email the client immediately, not even from the dashboard. The optional message is recorded in the request's audit trail and is not sent to the client. It refuses a request with no rejected file.

get_request

Where a Document Request stands, in one call: status, the checklist, every collected file with its verdict (approved, rejected with the reviewer's reason, or pending), and the reminder state. Give a request_id, or a search term matching the title or the client's name or email. When several requests match, the tool returns the candidates and asks for the id.

download_documents

Every collected file of a request, as one zip archive. The archive comes back embedded in the tool result as binary content (an MCP resource with a base64 blob and application/zip), not as a link: the API has no short-link endpoint for a zip, and the connector writes nothing to disk. What the agent does with the bytes is decided on the professional's side, exactly like a download from the browser. Large archives make large results; check with get_request that documents have arrived before calling it.

What the connector cannot do

Approving or rejecting a document is your decision, taken in the DokuTrak dashboard. No tool here can take it, and the service refuses it to any Agent Connection regardless of which connector asks. The same goes for billing, workspace settings and the management of API keys.

Revoking the Agent Connection in DokuTrak takes effect on the very next call: the connector answers with the service's 401 and nothing else.

Privacy Policy

The connector runs on your machine and talks to one service only: the DokuTrak API (https://app.dokutrak.com/api, or the URL you set in DOKUTRAK_API_URL), with the Agent Connection you configured. It sends what the tools need (the request you create, the id or search term you look up) and relays the answers to your agent. It writes nothing to disk, keeps no cache or log, and sends no telemetry or analytics to DokuTrak or to anyone else.

Everything it sends is processed by DokuTrak under the DokuTrak Privacy Policy, which covers what is collected, how it is used and stored, the subprocessors it is shared with, retention and deletion, and your rights. Questions: [email protected]; security issues: [email protected].

Development

npm cinpm run check   # typecheck, build, testsnpm test        # tests alone

The tests are contract tests at the MCP seam: a real MCP client and the real server, connected in memory through the official SDK's transport, with HTTP stubbed at fetch using recorded responses. They call tools, never functions, and run with no DokuTrak account and no network.

The staging run

Before a release, the built binary is driven once against a real workspace, by a real MCP client over stdio: create → chase → read → collect → revoke → 401. It is a record pasted into the release PR, never a CI check (a blocking check calls no third party). It pauses twice for acts the service refuses to any Agent Connection: rejecting the uploaded file, and revoking the key.

npm run buildDOKUTRAK_API_KEY=dk_live_… [email protected] npm run staging

A real Document Request is created and a real email goes to STAGING_RECIPIENT_EMAIL. The transcript lands in staging-run-<timestamp>.md (git-ignored); the key is never written to it.

The MCP Bundle

npm run bundle builds mcpb/server/index.cjs (the connector with every dependency inlined), starts it and checks that it lists exactly the tools declared in mcpb/manifest.json at the version of package.json, then validates the manifest and packs dokutrak.mcpb with the official mcpb CLI. CI runs it on every pull request, and the release workflow attaches dokutrak.mcpb to the GitHub release of each tag. A change to the tool surface or the version updates mcpb/manifest.json too.

Releasing

A tag vX.Y.Z matching package.json and SERVER_VERSION triggers .github/workflows/release.yml: npm run check, npm publish (trusted publishing through GitHub's OIDC token, provenance attached), then the listing in the MCP Registry as io.github.Crackx17/dokutrak-mcp — the mcpName of package.json, which the registry checks against the published tarball. Running the workflow by hand does a --dry-run and publishes nothing.

License

MIT.