Set up WorkerKit for Claude Duet.

Before you start

What you need.

A Claude Pro or Max plan on a US Claude account, admitted to the Claude Duet research preview from its waitlist. Steps 1 to 4 work without Duet, in claude.ai, Claude Desktop and Cowork.

A WorkerKit account, and admin access on it: choosing the key a connection will hold is an admin-only decision. The Free plan is enough, and no card is needed.

claude.ai in a browser, where the two connectors and the skill are added for Duet to pick up.

About a minute of the person's time. The agent does everything except approving the connection.

Why this route: Anthropic's help center gives Claude Duet no connector settings of its own. It says the memory, skills and connectors set up in Claude are there from the first message. So WorkerKit goes where Claude keeps them, as two custom connectors on claude.ai over the hosted WorkerKit MCP mounts, and the fleet connector signs in through WorkerKit's own OAuth server.

Step 5 confirms the connectors from inside Duet, and step 8 adds a second route into Duet, through the inbox it already reads.

The case for connecting at all is on the Claude Duet page.

Step 0. Human, under a minute

Add the public catalog.

Add the catalog mount on claude.ai with no sign-in. It proves the wiring before any account is involved.

• claude.ai: Customize > Connectors + > Add custom connector Name: WorkerKit Directory URL: https://mcp.workerkit.ai/directory Authentication: No sign-in

The catalog mount takes no credential at all, on either path. It is the fastest proof that the client can reach WorkerKit, and it is worth adding first for exactly that reason: if it connects and the fleet mount does not, the problem is the credential rather than the wiring.

Step 1. Human, about 60 seconds

Connect your fleet.

Add the fleet mount by URL and approve on the WorkerKit consent page. The reach is exactly the scopes on the key an admin picks there.

The fleet mount is an OAuth 2.1 authorization server as well as an MCP server, so it is added like the catalog: paste the URL and approve. No key is copied, and none is typed into Duet.

1. 01

   Give the client the URL and nothing else

   The fleet mount publishes OAuth discovery, so a client that supports it finds the authorization server on its own and registers itself. There is no client ID to create, no secret to store and no redirect address to register in advance.

2. 02

   Approve on the WorkerKit consent page

   The client sends the person to https://workerkit.ai/mcpauth. Signing in there is what proves who is connecting; nothing is typed into the assistant.

3. 03

   An account admin picks the key the client will hold

   The consent page lists the account's manager keys and can create one on the spot, so a first connect needs no visit to Fleet access beforehand. It is admin-only on the server, not just in the page: a member sees an honest notice rather than a broken screen.

4. 04

   Scopes are chosen with the key, and are the whole of the reach

   Whatever that key carries is exactly what the assistant can do, and nothing on it reaches a connected app. Narrow it now or widen it later; a new key drops into the same connection.

• claude.ai: Customize > Connectors + > Add custom connector Name: WorkerKit Workers URL: https://mcp.workerkit.ai/workers Then approve at https://workerkit.ai/mcpauth

Nothing is pasted into the chat on this path, which is its real advantage: the credential never passes through a message, a log or a screenshot, and revoking it later is one action in the dashboard rather than an edit inside the assistant.

A key carries the scopes it was granted when it was created. Granting a scope later never reaches an existing key: if a call answers 403 OPERATION_NOT_ALLOWED, an admin re-scopes the key at Fleet access or creates a new one, and the new key drops into the same place. Read key_info rather than assuming.

Step 2. Agent

Verify the scopes.

Call key_info on the fleet mount, from a Claude chat or from Duet. The scope list it returns is the truth, whatever was intended.

• Ask the key what it can do (no arguments) key_info

It returns the account title, the key name, the key's scopes and expiry, and serverTimeUtc. If a scope is missing it was either not granted or granted after the key was created, and an admin re-scopes it or creates a new key at Fleet access.

Step 3. Human and agent

Keep the WorkerKit skill.

Upload the skill in Claude, where Duet picks up skills from, or re-read the URL at the start of any thread that touches the fleet.

• claude.ai: Customize > Skills, upload a ZIP of workerkit-fleet/SKILL.md Source: https://workerkit.ai/claude-duet/SKILL.md

The file is a standard skill: a name and a description in its front matter, then the tool reference, the cost-field semantics and the safety rules. Uploading skills needs code execution and file creation turned on in Claude. Without the upload nothing is lost: the agent fetches the same URL before its first fleet call, and it is always the current version.

Step 4. Human, a minute

Check that Duet can see WorkerKit.

Ask Duet which connectors it can use, and confirm WorkerKit is among them before the smoke test.

• Paste into Duet's main thread Which connectors can you use right now? Is WorkerKit Workers among them? If it is, call key_info and tell me which account and scopes you have. Do not run or change anything.

If Duet lists WorkerKit and answers key_info, the fleet is connected and steps 6 and 7 can run from Duet. If it does not, run them from a Claude chat instead, where the same connectors answer, and use step 8 to bring results into Duet.

Step 5. Agent

Run the smoke test.

List the fleet and read recent runs, which alone prove the setup. If the account has workers, read one, run it with approval, and report the receipt.

If the fleet is empty, stop after step 3 and say so: a new account has no workers yet, and the steps above have already proved the setup, because the key authenticated and both read scopes answered. Do not read it as a fault, and do not invent a worker to run. Go to the first-worker step and give the fleet something to do.

If the fleet has workers, pick one whose effects are reads. Judge it from the instruction and the permissions, never from the name: a worker whose apps are all read-level and whose instruction only gathers, summarises or reports is safe to run; one that sends, posts, writes, orders or pays is not, whatever it is called. If none is clearly read-only, run one with an explicit prompt for that run bounding it to a read.

• 1. The directory is alive: confirm a kit count comes back directory_overview
• 2. The fleet list: report how many workers the assistant can see workers_list
• 3. Recent runs across the fleet: confirms readRuns runs_feed { "limit": 5 }
• 4. Before running anything, read what that worker does and what it may touch. Both ride on readWorkers, and this is the step that decides whether the run is safe instruction_get worker_permissions_get
• 5. Run it with the person's approval, then poll every 3 to 5 seconds until a terminal event arrives: run.succeeded, run.failed, run.blocked or run.awaiting_input worker_run run_events { "afterSeq": <lastSeq> }
• 6. Read the receipt: costAuthority first, then modelCostUsd run_get

Read costAuthority first, then modelCostUsd beside it, and report both. A null finalDigest is not a failure: a run produces one only when its harness wrote a report. A costAuthority of unbilled_estimate, none or in_flight is not a failure either; it says how far to trust the number, and the operating rules below say what each value means.

A Skipped run is also a pass, and it is the cost brake working: the gate refused the run before any money moved. Report the skipReason rather than retrying. Then report the fleet size, the active scopes, and what the test cost, and the fleet is live.

Step 6. Human and agent

Give the fleet its first worker.

A new account starts empty. Install a kit from its page in one click, or let the agent preview and install one once the person approves.

A new WorkerKit account starts with no workers, so the last step of setup is getting one. A kit is a ready-made worker: its instruction, its exact app permissions and its schedule, written and reviewed in advance, so installing one is how a fleet usually starts rather than a shortcut. There are two paths, and either ends with something to operate.

1. 01

   The one-click path: install a kit from its page

   Browse the catalog at workerkit.ai/kits, open a kit whose job you recognise, and install it. It arrives pre-instructed, pre-permissioned and pre-scheduled, and asks for whatever it needs to connect. This path needs no scope on the manager key at all, so it works on the first session whatever the key can do.

2. 02

   The agent path: shortlist, preview, approve, install

   With installKits on the key, the assistant can do the whole thing in chat: shortlist with kits_search, read the candidate with kit_get, then call kit_install_preview to surface its setup questions and exactly what the new worker would be allowed to touch. Put that in front of the person in plain words, and call kit_install once they approve, with deploy set to true so the new worker is actually able to run (that half needs manageDeployments on the key; without it the worker is installed but inert until worker_deploy). The new worker's key is shown once in the response.

3. 03

   Then run it once and read the receipt

   A worker earns its place on the first run. Trigger it with worker_run, follow run_events until a terminal event arrives, and report the digest and the settled cost from run_get. Set a schedule with schedule_create when the person is happy with what it did, and the fleet is not just connected, it is working.

Step 7. Human, a few minutes

Send reports to the inbox Duet reads.

Point a worker's email notification at the mailbox Duet briefs you from, so its result also reaches Duet in the morning brief.

1. 01

   Connect a sending mailbox in WorkerKit

   Under Connected apps in WorkerKit, connect a Google or Microsoft mailbox for sending notifications. It can be a different account from the one Duet reads.

2. 02

   Add an Email notification on the worker

   Open the worker, then Notifications, add Email, and enter the address of the Gmail account connected to Claude. Choose whether it reports on every run or only when something needs you.

3. 03

   Run it and find the report

   Run the worker once, confirm the delivery status in WorkerKit, then ask Duet for the latest WorkerKit report by sender. Duet checks at intervals, so the morning brief may be where it first appears.

Open Connected apps for the sending mailbox, and your workers to set the schedule once the first report reads well.

Classification on demand

Create a decision worker from your questions.

Categorize, score or triage connected data, then reuse and adjust the worker.

Reuse an existing worker or decision kit when its questions fit. For a custom classification task, creation compiles a supported source recipe and typed questions into a private kit and an installed decision worker. Optional deploy:true adds deployment; creation never starts a run or adds a schedule. The kit can be edited and published later through the normal kit lifecycle.

On the workers connection, call kit_app_tools(purpose:"decision") for recipes, argument schemas, permissions and examples, then kit_authoring_guide(section:"decision"). Check account connections with apps_list; public discovery does not check them. Send the request to decision_worker_create, follow nextCall, then use worker_run and run_get. Running needs runWorkers; results need readRuns. Use instruction_get / instruction_set for saved category or level answers, or answers on one run. Structural question/source changes need a revised kit and a replacement install. Check receipt status, coverage, omissions, warnings and cost before claiming completion.

Start with email-previews, calendar-events or sheets-rows. These recipes return judgments without app writes. Email previews do not include full threads; calendar events are invitation data, not transcripts; Sheets needs a Google file ID, a finite tab-qualified range and a columns map. A connected app alone does not make every tool a supported source. Source filters select evidence, not permissions: the Sheets recipe grants spreadsheet reads across the linked Drive account.

Supply 1–8 questions: choice with named options, score with ordered levels, or noul for the probability of a statement. Choice adds unclear automatically. Set an explicit confidenceFloor between 0 and 1; it routes uncertainty and does not promise accuracy. maxItems is 1–50, default 20. Creation requires createWorkers; optional deployment also needs manageDeployments.

Use a fresh requestId for each new worker. Retry with the same ID and identical body to recover the original receipt; a changed body returns 409. The response includes tokenId for MCP/CLI worker commands, stable workerId, kitSlug, readiness and nextCall. A deploymentError means the worker already exists: fix deployment on that worker. The receipt is a snapshot; check current worker readiness before running. No worker API secret is returned.

The workers connection includes discovery and authoring, so classification needs one MCP connection. Its full profile includes fleet management, app resource discovery, kit authoring and public directory reads. For a smaller list, a client that supports custom headers can send X-WorkerKit-Profile: decision on every request, including initialization, listing and calls, to select 24 workflow tools. A server operator can instead set MCP_WORKERS_PROFILE=decision. The endpoint and existing sign-in stay the same; a profile grants no additional permissions. Reconnect and relist after changing profiles. Use the advertised tool list: fleet administration, publishing, Action worker and event-stream tools require the full profile. Successful responses include structured content and a text fallback; creation declares an output schema.

Read the request example and full contract.

App tools, no model

Call app tools through an Action worker.

One exact app operation per call, with the inputs you supply.

An Action worker (modelType: "none") exposes 1 to 64 selected app tools and has no model, instruction, schedule or deployment. Each call invokes one tool with the arguments the caller sends, under the worker's own app access, and readiness.canInvoke on worker_get says whether its tools can be called now. Run and deploy requests answer 400 action_worker_requires_invoke.

On the workers connection's full profile: action_worker_options lists apps, tool schemas and defaults (readWorkers); action_worker_create copies an app's default tools, or takes an actionSpec of 1 to 64, plus a slackConnectionId for Slack tools when the account has more than one Slack bot (createWorkers); worker_tools lists the saved tools (readWorkers); worker_tool_call calls one with arguments, plus an idempotencyKey on a write (invokeActions); worker_tool_invocation reads a write's receipt (invokeActions). Pass the returned tokenId. kits_search(modelType:"none") finds Action kits; install one with kit_install and no deploy.

A write tool (readOnly: false) needs an idempotency key. Repeating the key with the same arguments returns the stored receipt (a confirmation, not the app's response) without calling the app again; with other arguments it is a 409. A call answers with status succeeded, denied, failed, in_progress or outcome_unknown; branch on it, not on the HTTP code. On outcome_unknown the write may have happened: never resend it automatically, under any key; read the receipt and check the app. Each call counts toward the worker's daily request limit; a refusal marked retryable, such as usage_limit, is retried later under a new key. Only the owner changes the selected tools, in the dashboard.

Read the call examples.

Least privilege

What each scope grants.

The key's scopes are exactly its reach, and a refusal names the scope it wanted. Start with the reads and add scopes as the work asks for them.

Operating rules

What the agent follows on every call.

The same rules the skill file carries, so a session that read neither still has them. Every row is a fact of the platform, not advice.

Quick answers

Symptoms, causes, fixes.

Each WorkerKit failure names itself in the response, so an agent never has to guess which one it met: read the code, apply the row, and carry on.