# WorkerKit > WorkerKit is building the cloud for the AI workforce: a private environment for each company, with controls for access and spending and the compute that runs the work. Connect workers to Claude, ChatGPT, Gemini, Microsoft Copilot or another supported assistant through MCP, the REST API or the CLI. Each worker has a job, its own app permissions and spending limits, and a record of its work. Workers retain their configuration and memory between runs, can run in the background on schedules or events, and can deliver important information to the right person in time to decide. Kits are reusable worker definitions; coordinating kits can be installed with supporting kits as a connected group of workers. Language workers can test instruction improvements against real runs, with approval or within configured limits. Browse kits at https://workerkit.ai/kits. Last updated: 2026-11-13. ## Docs - [About WorkerKit](https://workerkit.ai/about.md): the full product description in markdown for AI agents, with how it works, the access model, model choice, where workers run, complete pricing, both FAQs and the site map - [All kits](https://workerkit.ai/kits.md): every public kit in the directory, with the job each one does, the apps it needs and its permanent page URL, plus the category hubs. Regenerated from the live catalog on every deploy - [Pricing](https://workerkit.ai/pricing.md): the shared cloud plan of record in markdown, with prices, the yearly runtime credit, the per-plan limits table, what every plan includes, and how the wallet and the caps behave. - [Glossary](https://workerkit.ai/glossary.md): every defined term in full, one fetch for the vocabulary around AI workers and workforces (kits, workflows, skills, plugins, MCP, runtimes, schedules, tool calls, wallet, BYOK). Each term is also a page at https://workerkit.ai/glossary/{slug}. Regenerated on every deploy - [Blog](https://workerkit.ai/blog.md): an index of every post, with its description and dates. Each post page at https://workerkit.ai/blog/{slug} carries its full article in the initial HTML, readable without executing JavaScript. Regenerated on every deploy - [Guides](https://workerkit.ai/guides.md): every step-by-step guide, grouped by job, with each guide's own markdown address. Each guide is a page at https://workerkit.ai/guides/{slug}, and its full text (the kits it installs, the setup steps in order, its FAQ) is at https://workerkit.ai/guides/{slug}.md. Regenerated on every deploy - [What is an AI workforce?](https://workerkit.ai/glossary/ai-workforce): the roles of people, AI models, workers, apps and company controls, plus the distinction between a kit, workflow and workforce - [From AI tools to an AI workforce](https://workerkit.ai/blog/from-ai-tools-to-an-ai-workforce): how people, AI models and workers operate together in managed cloud infrastructure, with a support workflow design and ownership and acceptance checks - [Cost per acceptable AI result](https://workerkit.ai/blog/cost-per-acceptable-ai-result): an evaluation worksheet covering execution, failed attempts, review, correction and maintenance; a buying framework, not a metric WorkerKit computes - [AI workers for inbound sales](https://workerkit.ai/guides/ai-workers-for-inbound-sales.md): setup steps, app permissions, draft review and the human handoff between intake and deal follow-up; human page at https://workerkit.ai/guides/ai-workers-for-inbound-sales - [Enterprise AI automation](https://workerkit.ai/blog/enterprise-ai-automation): a practical blueprint with use cases, a reusable process brief, ownership, app authority and acceptance criteria - [Evaluating enterprise AI automation platforms](https://workerkit.ai/blog/enterprise-ai-automation-platform-evaluation): mandatory requirements, operational failure checks, a weighted preference worksheet and the evidence for a buying decision - [AI agents vs RPA](https://workerkit.ai/blog/ai-agents-vs-rpa): choose interpretation, application access and coordination per process step, with a support-intake example - [AI workers for support triage](https://workerkit.ai/guides/ai-workers-for-support-triage.md): setup, coverage checks, email drafts, task writes and the receiving team's responsibilities; human page at https://workerkit.ai/guides/ai-workers-for-support-triage - [API for agents](https://workerkit.ai/api.md): how a non-MCP AI platform connects to WorkerKit over plain HTTPS, with the two OpenAPI spec URLs, the API-key and OAuth2 paths, the scopes, a curl quick start, the operating rules and the error envelope - [Create a decision worker](https://workerkit.ai/api.md#decision-workers): classify email previews, calendar events or Google Sheets rows from typed questions. API POST /api/manage/decision-workers, MCP decision_worker_create, CLI wk decision create. Creates a private kit and worker; optional deploy never runs it. MCP supports a focused decision profile. - [Call an Action worker](https://workerkit.ai/api.md#action-workers): selected app tools with no model, called one at a time with exact inputs. Create with POST /api/manage/workers/actions (MCP action_worker_create, full profile, scope createWorkers), call with POST /api/manage/workers/{tokenId}/tools/{toolName}/invoke (MCP worker_tool_call, scope invokeActions). Never deployed or run. - [AI workers for Meta Muse](https://workerkit.ai/muse.md): discover and build workers for research, app tasks and custom classification. Muse operates the fleet through the HTTPS API with a scoped manager key. Human page: https://workerkit.ai/muse - [Set up WorkerKit in Meta Muse](https://workerkit.ai/muse/setup.md): the procedure for the Muse agent, step by step: create the manager key, add the custom connector, verify scopes with key-info, install the skill file, run the smoke test; then the operating rules and the troubleshooting table. Human page: https://workerkit.ai/muse/setup - [AI workers for Grok Bot](https://workerkit.ai/grok-bot.md): connect WorkerKit at https://mcp.workerkit.ai/workers with a scoped manager key to discover kits, build workers and operate your fleet. Human page: https://workerkit.ai/grok-bot - [Set up WorkerKit in Grok Bot](https://workerkit.ai/grok-bot/setup.md): connect WorkerKit at https://mcp.workerkit.ai/workers, verify access, read the skill, check deployment and funding, then run a job and inspect the result. Human page: https://workerkit.ai/grok-bot/setup - [Grok Bot skill file](https://workerkit.ai/grok-bot/SKILL.md): the skill a Grok Bot agent installs as workerkit/SKILL.md: the MCP tool reference for WorkerKit and the separate Knowledge MCP, the cost-field semantics and the safety rules for operating a fleet - [WorkerKit skill file](https://workerkit.ai/muse/SKILL.md): the skill an agent installs as skills/workerkit/SKILL.md: the endpoint reference, the cost-field semantics and the safety rules for operating a fleet over the API - [AI workers for Claude](https://workerkit.ai/claude.md): connect WorkerKit at https://mcp.workerkit.ai/workers through OAuth, or with an Authorization header in Claude Code, to discover kits and operate workers. Human page: https://workerkit.ai/claude - [Set up WorkerKit in Claude](https://workerkit.ai/claude/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/claude/SKILL.md. Human page: https://workerkit.ai/claude/setup - [AI workers for ChatGPT](https://workerkit.ai/chatgpt.md): connect WorkerKit at https://mcp.workerkit.ai/workers through developer mode and OAuth sign-in to discover kits and operate workers. No pasted key needed. Human page: https://workerkit.ai/chatgpt - [Set up WorkerKit in ChatGPT](https://workerkit.ai/chatgpt/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/chatgpt/SKILL.md. Human page: https://workerkit.ai/chatgpt/setup - [AI workers for Gemini](https://workerkit.ai/gemini.md): why Gemini gets a workforce of background AI workers from WorkerKit, and how the two connect: which surface decides the door: Antigravity takes an Authorization header in its MCP config, and the consumer Gemini app connects a custom app by URL through the sign-in flow. Human page: https://workerkit.ai/gemini - [Set up WorkerKit in Gemini](https://workerkit.ai/gemini/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/gemini/SKILL.md. Human page: https://workerkit.ai/gemini/setup - [AI workers for Microsoft Copilot](https://workerkit.ai/microsoft-copilot.md): why Microsoft Copilot gets a workforce of background AI workers from WorkerKit, and how the two connect: a Copilot Studio agent's Model Context Protocol tool, then published to the Teams and Microsoft 365 Copilot channel. There is no field inside Microsoft 365 Copilot itself. Human page: https://workerkit.ai/microsoft-copilot - [Set up WorkerKit in Microsoft Copilot](https://workerkit.ai/microsoft-copilot/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/microsoft-copilot/SKILL.md. Human page: https://workerkit.ai/microsoft-copilot/setup - [AI workers for OpenClaw](https://workerkit.ai/openclaw.md): why OpenClaw gets a workforce of background AI workers from WorkerKit, and how the two connect: an mcp.servers entry in the self-hosted Gateway's config, with an Authorization header. The simplest setup of the seven. Human page: https://workerkit.ai/openclaw - [Set up WorkerKit in OpenClaw](https://workerkit.ai/openclaw/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/openclaw/SKILL.md. Human page: https://workerkit.ai/openclaw/setup - [AI workers for Hermes](https://workerkit.ai/hermes.md): why Hermes gets a workforce of background AI workers from WorkerKit, and how the two connect: an mcp_servers entry in ~/.hermes/config.yaml with an Authorization header, and a skill installed in one command from a URL. Human page: https://workerkit.ai/hermes - [Set up WorkerKit in Hermes](https://workerkit.ai/hermes/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/hermes/SKILL.md. Human page: https://workerkit.ai/hermes/setup - [AI workers for Instinct](https://workerkit.ai/instinct.md): why Instinct gets a workforce of background AI workers from WorkerKit, and how the two connect: the honest answer: as of 2026-09-15 Instinct documents no MCP server, connector or inbound API, so there is no connection to make. What does work is a worker delivering its run notification to the assistant's own email address. Human page: https://workerkit.ai/instinct - [Set up WorkerKit in Instinct](https://workerkit.ai/instinct/setup.md): the procedure, written for the agent doing it, step by step, then the scopes, the operating rules and the symptom-to-fix table. The skill file it keeps: https://workerkit.ai/instinct/SKILL.md. Human page: https://workerkit.ai/instinct/setup - [Give your dot a workforce](https://workerkit.ai/dots.md): discover kits or build custom workers for research, app tasks, monitoring and classification. Connect apps and custom MCP servers, clone workers, coordinate runs and bring back results. Includes an agent starting workflow. Human page: https://workerkit.ai/dots - [Set up WorkerKit with your dot](https://workerkit.ai/dots/setup.md): kit discovery and account access through OAuth, first-run readiness and email reports. Skill reference: https://workerkit.ai/dots/SKILL.md. Human page: https://workerkit.ai/dots/setup - [OpenAI Dots explained](https://workerkit.ai/blog/openai-dots-explained): launch-day features, availability, controls and how a worker fits. - [OpenAI Dots vs Meta Muse vs Grok Bot](https://workerkit.ai/blog/dots-vs-muse-vs-grok-bot): a source-linked comparison of computers, channels, connected apps and controls. - [AI workers for Hark](https://workerkit.ai/hark.md): why Hark gets a workforce of background AI workers from WorkerKit, and how the two connect: WorkerKit at https://mcp.workerkit.ai/workers with a narrow manager key, the route tested on 2026-10-06. Human page: https://workerkit.ai/hark - [Set up WorkerKit in Hark](https://workerkit.ai/hark/setup.md): the procedure, written for the agent doing it, step by step, including how to keep a key Hark asks for in the chat narrow and out of its memory. The skill file it keeps: https://workerkit.ai/hark/SKILL.md. Human page: https://workerkit.ai/hark/setup - [Hark Pro vs Meta Muse vs Instinct vs Grok Bot](https://workerkit.ai/blog/hark-vs-muse-vs-instinct-vs-grok-bot): a source-linked launch-day comparison of computers, memory, approvals, pricing and connections. - [AI workers for Claude Duet](https://workerkit.ai/claude-duet.md): add WorkerKit at https://mcp.workerkit.ai/workers as a custom connector on claude.ai, then check that Duet can use it. Workers can also email reports to Duet's inbox. Human page: https://workerkit.ai/claude-duet - [Set up WorkerKit for Claude Duet](https://workerkit.ai/claude-duet/setup.md): the procedure, written for the agent doing it: the WorkerKit connector and skill on claude.ai, signed in over OAuth, a check from Duet itself, a smoke test and reports to the inbox Duet reads. The skill file it keeps: https://workerkit.ai/claude-duet/SKILL.md. Human page: https://workerkit.ai/claude-duet/setup - [What is Claude Duet?](https://workerkit.ai/blog/claude-duet-explained): launch-day features, access, usage limits, approvals, stated limitations and how a worker fits. - [Claude Duet vs OpenAI Dots vs Meta Muse](https://workerkit.ai/blog/claude-duet-vs-dots-vs-muse): a source-linked comparison of computers, channels, work while away, approvals, access, cost and how each reaches a fleet. - [AI workers for Lovable](https://workerkit.ai/lovable.md): add WorkerKit at https://mcp.workerkit.ai/workers on Lovable's Connectors page with OAuth. Published apps use a REST API connector with a read-only key to show worker results. Human page: https://workerkit.ai/lovable - [Set up WorkerKit in Lovable](https://workerkit.ai/lovable/setup.md): the procedure, written for the agent doing it: the WorkerKit server, the OAuth approval, the scope check, the skill as a workspace skill and in workspace knowledge, a smoke test, and the optional custom connector. The skill file it keeps: https://workerkit.ai/lovable/SKILL.md. Human page: https://workerkit.ai/lovable/setup - [OpenAI Decisions API: the public beta explained](https://workerkit.ai/blog/openai-decisions-api-public-beta): what launched on 2026-10-06, the three question types, pricing, how it differs from Jev, and how a WorkerKit decision worker uses either classifier. - [OpenAI Decisions API vs Jev, field by field](https://workerkit.ai/blog/openai-decisions-api-vs-jev): a source-linked comparison of the two published contracts: question types, confidence, refusals, images, limits, data terms and price. - [Developers hub](https://workerkit.ai/developers.md): the four doors of the developer surface, with the one line that opens each: the MCP server, the wk CLI, the open-source packages and the plain-HTTPS API. Generated from the page on every deploy - [The WorkerKit MCP server](https://workerkit.ai/mcp.md): connection steps, authentication and tool reference for WorkerKit at https://mcp.workerkit.ai/workers and the separate Knowledge MCP at https://mcp.workerkit.ai/knowledge. Generated from the page on every deploy. Human page: https://workerkit.ai/mcp - [The WorkerKit CLI](https://workerkit.ai/cli.md): what the wk command covers by area, how to install it from npm or Homebrew, and how its browser-approved sign-in works. Generated from the page on every deploy. Human page: https://workerkit.ai/cli - [WorkerKit on npm](https://workerkit.ai/npm.md): the MIT-licensed packages and repositories the CLI and the MCP server are both generated from. Generated from the page on every deploy. Human page: https://workerkit.ai/npm - [Partner programs](https://workerkit.ai/partners.md): the two partner tracks, what acceptance brings, the three-step process and the partner FAQ. Generated from the page on every deploy. Human page: https://workerkit.ai/partners - [For Publishers](https://workerkit.ai/publish.md): who the directory is written for, the three moves of building a kit, what a kit page earns its publisher, the three ways a kit ships and the publisher FAQ. Generated from the page on every deploy. Human page: https://workerkit.ai/publish - [First Workflow Program](https://workerkit.ai/first-workflow.md): a company's first AI workflow, built with WorkerKit engineers: a free 45-minute assessment with a written recommendation, then for selected companies the workflow built, tested and launched in their apps, implementation included; what is included, what they keep and the program FAQ. Generated from the page on every deploy. Human page, with the application form: https://workerkit.ai/first-workflow - [Build a kit](https://workerkit.ai/build.md): what a kit carries, the six steps every surface shares, the three ways in (an AI assistant over MCP, the API, the wk CLI), a complete example kit, publishing a worker you already tuned, the rules every publish enforces, and managing a kit after it ships. Generated from the page on every deploy. Human page: https://workerkit.ai/build - [Build a kit with your AI assistant](https://workerkit.ai/build/mcp.md): connecting an MCP client, the scopes a build uses, what to ask, and the tool sequence the assistant runs from the guide to a public listing. Generated from the page on every deploy. Human page: https://workerkit.ai/build/mcp - [Build a kit over the API](https://workerkit.ai/build/api.md): every request from the authoring guide to a public listing, as curl, and the responses to branch on. Generated from the page on every deploy. Human page: https://workerkit.ai/build/api - [Build a kit from the terminal](https://workerkit.ai/build/cli.md): the wk commands from sign-in to promotion, and how to run them unattended. Generated from the page on every deploy. Human page: https://workerkit.ai/build/cli - [Kit-authoring skill](https://workerkit.ai/build/SKILL.md): the procedure, the publish rules and the quality bar, written for an agent to keep, with the MCP tool, the REST request and the wk command for every step - [What WorkerKit is, for agents](https://api.workerkit.ai/api/directory/mcp/about): the same text the MCP server's workerkit_about tool serves, no auth, one section at a time (`?section=index|why|operate|access|cost|start`): the request shapes that call for a worker, how a fleet is operated from an agent's seat, the access model, the money rules, and how to connect from an MCP host, a REST host or a terminal ## What WorkerKit is - **Job:** the work to be done. - **Worker:** a configured agent that carries out a job in your account, with its own access, limits and record. - **Workforce:** the workers a company or team runs. They do not all need to be connected. - **Workflow:** the process workers carry out together. - **Kit:** a reusable worker definition containing instructions or decision rules, app permissions and run settings. A coordinating kit can be installed with supporting kits to create a connected group of workers. Language workers follow instructions; decision workers evaluate defined questions and return structured judgments. An Action worker has no model and exposes selected app tools for a caller to invoke with exact inputs. Each works within the app access you grant it. Installing a kit creates a worker in your own account; the kit is the definition, and the worker is the configured instance. Workers can run on demand, on a schedule or when an event arrives. Scheduled and event-driven work can watch connected apps, hand work to other workers and deliver a report or decision request without someone asking for each update. WorkerKit hosts those runs, so your laptop and assistant do not have to remain open. Persistence means the worker's configuration, memory and run history remain between runs; it does not mean continuous model execution. WorkerKit is in early access. Browse kits and agent setup guides now. Plan your environment at https://workerkit.ai/#request: say what your AI workforce should run so we can plan your workers, app connections, access and spending limits, and discuss availability. Published plans and prices apply to shared cloud. Or join the waitlist at https://workerkit.ai/waitlist; once invited, connect your agent and apps, choose a kit, review its access and limits, and run your first worker. To start with one process, apply for a free cloud workforce assessment at https://workerkit.ai/first-workflow. The public kit marketplace at https://workerkit.ai/kits is free to browse without an account. Language kits provide a starting model that you can change; compare supported models on your own examples. Model tokens for hosted runs are billed at the provider list price through a prepaid wallet or your own provider key. Running workers on your own infrastructure is coming later. ## Why WorkerKit - Shared workers for every AI you use: authorized assistants reach the same workforce over MCP, the REST API or the CLI. Each worker keeps its configuration and memory when you switch between supported assistants. - Start with one job: once invited, choose a kit, connect your agent and apps, review access and limits, and run your first worker. - Limited app access, set per worker: access is granted per app and per level (Off / Read / Write), can be narrowed or revoked at any time, and the worker keeps running with whatever is left. A worker cannot reach an account you have not connected. - The safety layer is not an upsell: the app firewall, redaction and contact rules ship on every plan, Free included. - Choose a model for the job: language kits provide a starting model. Compare supported models on your own examples and change the model when another gives a better balance of cost and quality. - Compare models on your own work: send a run to a different model and review its receipt before choosing a default. - No markup on model tokens: WorkerKit hosts the runs and the tokens bill at the provider list price to the cent, from a prepaid wallet or from your own provider key (BYOK: first 1M tokens a month free of platform fee, then 5% of list). The platform's own fees are the plan price and the wallet top-up fee (5.5% with a $1 minimum), both stated in the open. - Background work that reaches people: scheduled or event-driven workers watch connected apps, hand work to each other and deliver reports or decision requests without a prompt for every update. Every run leaves a record. - Improvement within your limits: a language worker can check results against its configured quality criteria and test instruction changes against real runs. Changes wait for approval or apply within the limits you set after passing comparison checks. Applying a change updates instructions for future runs; memory is retained separately under the worker's memory settings. - Published plans and prices apply to shared cloud. - A free plan that is real: 5 workers, 500 tool calls a day and the complete safety layer, with no card and no expiry. - Invite teammates on any plan, each with their own normal free usage once. Free and Pro include one execution seat; Team is $299 a month flat with ten execution seats. Account limits still apply; only an Admin can upgrade or assign seats. - A public kit marketplace built in: discover kits others published, publish your own from your AI assistant, the API or the CLI, with a permanent address for life, and keep kits private when you want to; private kits are included on every plan. ## Model types Browse both kinds at https://workerkit.ai/kits. Decision model kits return typed judgments and route work through defined rules: https://workerkit.ai/kits/decision-models. Language model kits follow instructions to research, write, and act through permitted tools: https://workerkit.ai/kits/language-models. An AI agent can run a decision worker and use the returned references and answers to continue a task. Language-provider choices and transcripts apply to language workers; decision workers use their configured decision model and structured results. Action workers have no model: they expose selected app tools that an agent calls with exact inputs, and they are never deployed or run (https://workerkit.ai/api.md#action-workers). ## Key pages - [Home](https://workerkit.ai/): the cloud for the AI workforce: how AI supplies the intelligence, WorkerKit runs it and the company keeps the authority; the worker types; the company's isolated environment (access, app gateway, observability, one source of truth, work knowledge base, spending limits, secrets vault, network rules and self-improvement); kits by team; the environment request at https://workerkit.ai/#request; and the buyer FAQ - [Kit directory](https://workerkit.ai/kits): every public kit, searchable by job and filterable by app, model and how it runs - [Pricing](https://workerkit.ai/pricing): shared cloud plans with Free $0; Pro $29 per month; Team $299 per month flat with ten execution seats; Enterprise custom; invites on any plan; yearly plans load a runtime credit into your wallet on day one ($50 Pro, $500 Team). - [Build a kit](https://workerkit.ai/build): how to build and publish a kit from an AI assistant over MCP, the API or the wk CLI, with one page per surface and a skill file for agents at https://workerkit.ai/build/SKILL.md - Kit pages: https://workerkit.ai/kit/{slug}, one permanent page per kit with its job, apps and access. Markdown summaries are at https://workerkit.ai/kit/{slug}.md; every published kit is listed at https://workerkit.ai/kits.md - Category hubs: https://workerkit.ai/for/{category}, one page per job or role, indexed at https://workerkit.ai/for-jobs and https://workerkit.ai/for-roles - App hubs: https://workerkit.ai/for-apps/{app}, one page per app, all of them listed at https://workerkit.ai/for-apps - [Guides](https://workerkit.ai/guides): step-by-step guides to handing one job to AI workers, each built on real kits, grouped by job - [Glossary](https://workerkit.ai/glossary): plain definitions for the vocabulary around AI workers, one page per term - [Blog](https://workerkit.ai/blog): long-form writing on how AI workers are built, run and paid for - [Developers](https://workerkit.ai/developers): the developer surface's hub, with one page per product: the MCP server at https://workerkit.ai/mcp, the plain-HTTPS API at https://workerkit.ai/api, the wk CLI at https://workerkit.ai/cli, and the open-source npm packages at https://workerkit.ai/npm - [Partner programs](https://workerkit.ai/partners): the two partner tracks, with an application form: technology partners make their product connectable by AI workers, solutions partners build and run workforces for their clients - [For Publishers](https://workerkit.ai/publish): written for professionals and experts in their field (practitioners, consultants and agencies, specialists in a technical or regulated practice), covering how publishing works: building from an AI assistant, the API or the CLI, public, private and protected kits, and the permanent page every public kit earns - [First Workflow Program](https://workerkit.ai/first-workflow): apply for a free workflow assessment; selected companies get their first AI workflow built with WorkerKit engineers, implementation included, and keep it either way. The application form books the assessment call - [Terms of service](https://workerkit.ai/terms), [Privacy](https://workerkit.ai/privacy), [Acceptable use](https://workerkit.ai/aup), [Security](https://workerkit.ai/security) ## Developers - Connect **WorkerKit** at https://mcp.workerkit.ai/workers to discover kits, build workers and operate your fleet. Sign in through OAuth or use a scoped, revocable bearer key. The separate Knowledge MCP at https://mcp.workerkit.ai/knowledge searches and reads published worker reports with the same sign-in or key. Workers use their own app grants; readRuns and invokeActions can return data from those apps. Profiles, tools and setup: https://workerkit.ai/mcp - The plain-HTTPS API, for platforms that cannot speak MCP (custom connectors, ChatGPT Actions, Gemini extensions): the same fleet at https://api.workerkit.ai/api/manage/workers behind an `Authorization: Bearer pe_mgr_...` key or OAuth2 with PKCE, and the same public directory at https://api.workerkit.ai/api/directory/mcp with no auth. OpenAPI specs: https://api.workerkit.ai/openapi/workers.json and https://api.workerkit.ai/openapi/directory.json. Page: https://workerkit.ai/api (markdown: https://workerkit.ai/api.md) - Grok Bot speaks MCP: connect WorkerKit at https://mcp.workerkit.ai/workers with a scoped manager key in protected configuration. The connection includes kit discovery. The case is at https://workerkit.ai/grok-bot (markdown: https://workerkit.ai/grok-bot.md), setup at https://workerkit.ai/grok-bot/setup (markdown: https://workerkit.ai/grok-bot/setup.md), and the skill at https://workerkit.ai/grok-bot/SKILL.md - Meta Muse, Meta's personal AI agent, connects over the plain-HTTPS API as a custom connector with a manager key, since it adds outside services over their APIs rather than as MCP servers. The case is at https://workerkit.ai/muse (markdown: https://workerkit.ai/muse.md), the procedure for the Muse agent at https://workerkit.ai/muse/setup (markdown: https://workerkit.ai/muse/setup.md), and the skill file it installs at https://workerkit.ai/muse/SKILL.md. Muse Code, Meta's terminal coding agent, is a different product and speaks MCP: https://workerkit.ai/blog/muse-code-and-ai-workers - OpenAI Dots uses WorkerKit at https://mcp.workerkit.ai/workers with OAuth for kit discovery and fleet operations. Workers can also deliver reports to the dot's inbox. The case: https://workerkit.ai/dots (markdown: https://workerkit.ai/dots.md); setup: https://workerkit.ai/dots/setup (markdown: https://workerkit.ai/dots/setup.md); skill reference: https://workerkit.ai/dots/SKILL.md. - Hark speaks MCP: it connects to https://mcp.workerkit.ai/workers with a scoped manager key in the Authorization header (tested 2026-10-06). Hark asks for the key in the chat, so the setup makes it a dedicated, read-first, expiring key. The case: https://workerkit.ai/hark (markdown: https://workerkit.ai/hark.md); setup: https://workerkit.ai/hark/setup (markdown: https://workerkit.ai/hark/setup.md); skill: https://workerkit.ai/hark/SKILL.md. - Claude Duet uses the connectors set up in Claude. Add **WorkerKit** on claude.ai at https://mcp.workerkit.ai/workers with OAuth, then ask Duet to confirm it sees the connection. Workers can also email reports to Duet's inbox. The case: https://workerkit.ai/claude-duet (markdown: https://workerkit.ai/claude-duet.md); setup: https://workerkit.ai/claude-duet/setup (markdown: https://workerkit.ai/claude-duet/setup.md); skill: https://workerkit.ai/claude-duet/SKILL.md. - Lovable connects any remote MCP server as a chat connector, on every plan: WorkerKit at https://mcp.workerkit.ai/workers with OAuth (lovable.dev is on the fleet mount's registration allowlist). Chat connections are personal and never part of a published app. The case: https://workerkit.ai/lovable (markdown: https://workerkit.ai/lovable.md); setup: https://workerkit.ai/lovable/setup (markdown: https://workerkit.ai/lovable/setup.md); skill: https://workerkit.ai/lovable/SKILL.md. - The open-source wk CLI manages a worker fleet from the terminal — `wk about` (what WorkerKit is and when to use it, written for an agent), runs, schedules, memory, instructions, deliveries, budgets, cloning, kit authoring (`wk kit tools|guide|vocabulary|validate|publish`), connected apps (`wk apps`), model keys (`wk model-keys`) and the account's own MCP servers (`wk mcp-servers`): install with `npm install -g @workerkit/cli` or `brew install workerkit/tap/wk`. MIT licensed at https://github.com/workerkit/cli, published as https://www.npmjs.com/package/@workerkit/cli. The `--json` flag is a stable machine-readable contract for scripts and agents, and the kit-directory commands work without an account. Page: https://workerkit.ai/cli - The MCP server and the CLI are built from @workerkit/core (https://github.com/workerkit/core, https://www.npmjs.com/package/@workerkit/core), the MIT-licensed API client and tool-descriptor registry: one definition per tool, so the CLI and the MCP server cannot drift apart. The REST specs are generated by the same backend on a parity rule, so the three describe one surface. Page: https://workerkit.ai/npm - Building a kit: https://workerkit.ai/build covers all three surfaces, one page each (https://workerkit.ai/build/mcp, https://workerkit.ai/build/api, https://workerkit.ai/build/cli), and https://workerkit.ai/build/SKILL.md is the skill an agent keeps for it. - The hub: https://workerkit.ai/developers ## FAQ Connect through [MCP](https://workerkit.ai/mcp), the [API](https://workerkit.ai/api) or the [CLI](https://workerkit.ai/cli). See [kit building](https://workerkit.ai/build) for setup and [pricing](https://workerkit.ai/pricing) for shared cloud plans. **What is WorkerKit?** WorkerKit is building a managed cloud for a company’s AI workforce: workers, apps and databases in an isolated environment, with controls for access, credentials, connections and spending. People and AI assistants assign work; WorkerKit runs it and keeps the record. **How is WorkerKit built for the AI workforce?** WorkerKit is built from the ground up for a workforce where people and AI work together. People set direction. AI models coordinate. Workers carry out the work, connected to the apps, data and shared knowledge they need. Every worker has a defined role, explicit permissions and clear limits. WorkerKit provides the managed cloud for this workforce to run, collaborate and improve under your control. **Why use WorkerKit if my AI assistant already runs tasks?** WorkerKit lets your assistant create and run workers, apps and databases in your company’s isolated cloud environment, with the infrastructure hosted and managed for you. Workers can keep running when your laptop or chat is closed, including scheduled and event-driven jobs. Work that needs access to business systems runs with protected credentials, explicit app permissions, network rules, spending limits and activity records. Your team and supported AI assistants can share workers, and their instructions, approved memory and controls persist between jobs and when you change assistants. **What is the difference between a worker, a kit and a workflow?** A worker carries out a job with its own configuration, access and limits. A kit is a reusable worker definition: its instructions or decision rules, app permissions and run settings. Installing a kit creates a worker in your account. A workflow is a process workers carry out together, with defined handoffs between them. **What types of workers can run in my environment?** Language workers research, write and analyze. Decision workers classify, score and route work. Action workers make selected app calls without a model. Code workers use Codex or Claude Code; browser workers handle web tasks; app and DB workers run applications and databases. Open-source workers deploy and run projects from GitHub. **Can AI agents create and manage workers?** Yes. Connect a supported assistant, such as Claude or ChatGPT, through WorkerKit’s MCP server. Other clients can use the REST API or the wk CLI. Agents can discover kits and create, configure and run workers within the permissions you grant. OAuth or scoped, revocable keys control access. **Where do workers run, and can they work without me?** Workers run in your company’s environment in the WorkerKit managed cloud, so your laptop and AI assistant can be closed. Language and decision workers can run on demand, on schedules or from supported events. Action workers execute the app calls they receive. Connected delivery channels can bring reports and decision requests to your team. **Which AI models can I use, and can I switch them?** Language workers support models from OpenAI, Anthropic, Google and xAI; decision workers use Jev. You can change a language worker’s supported model while keeping its app permissions and spending limits. Action workers make app calls without a model. **What can a worker access or change?** You choose the connected apps and allowed actions for each worker. Built-in app controls can also restrict the fields it sees. Requests outside those permissions are blocked and recorded. Stopping a run does not undo an email, app update or other action already completed. **Do credentials or business data reach the AI model?** App credentials and model keys stay in WorkerKit’s vault; WorkerKit makes authenticated calls on the worker’s behalf. Business data the worker is allowed to read may be sent to its selected model to complete the job. Built-in app controls let you hide selected fields before they reach the model. **What does a worker remember between runs?** Worker instructions, permissions and settings persist between runs. Worker memory holds the rules and facts you give it or approve; proposed memory changes wait for your approval. Run history is retained according to your plan. You decide which context should carry into future work. **How does the work knowledge base work?** The work knowledge base makes published worker reports searchable. Account admins choose the report sources and readers. People, workers and connected AI agents can retrieve the reports their access allows, including through the separate Company Knowledge MCP connection. Report access does not grant access to the source apps or run transcripts. **How do workers improve over time?** When continuous learning is enabled, language workers check results against your quality criteria and can propose instruction changes. You can review changes yourself or let changes that pass comparison checks apply automatically within your limits. Learning does not grant extra app access or increase a worker’s spending limits. **How do I control costs and see what workers did?** Set spending limits per run, per day and across your account. Run records show the work performed, app calls, blocked requests and costs. Use those records to review outcomes and measure cost per acceptable result before expanding the work. **Who maintains the app connections?** WorkerKit maintains its built-in app integrations. Your team connects the accounts and grants each worker the access it needs. If an app login expires, reconnect that account. Your team or the provider maintains any custom MCP server you connect. **How do we get started? Do we need to code?** You can start without coding by choosing a ready-made kit or describing a custom job to a supported AI assistant. Request an environment and tell us the work, apps and data involved. We’ll help plan the workers, connect the required accounts and set access and spending limits. For one process, start with a free cloud workforce assessment. Explore [what an AI workforce is](https://workerkit.ai/glossary/ai-workforce) and [how people, models and workers work together](https://workerkit.ai/blog/from-ai-tools-to-an-ai-workforce). ## Notes for AI agents - Public product and editorial pages, catalog indexes, kit listings, category hubs and app hubs carry their own visible content and links in the initial HTML. Published kit descriptions and public setup information are included when available. The browser then loads the interactive page; account state and install checks use the live service. - Homepage kit cards and job/role category previews use public data from the site's latest build. These previews are dated snapshots, not live availability or account state. Open the kit page to check its current details before installing. The full directory remains searchable at https://workerkit.ai/kits. - Public pages carry JSON-LD: WebSite, Organization, ItemList and FAQPage on the directory; SoftwareApplication and BreadcrumbList on kit pages; CollectionPage, ItemList and BreadcrumbList on category and app hubs; BlogPosting on blog posts; TechArticle, HowTo and FAQPage on guides; DefinedTerm and DefinedTermSet on the glossary; WebAPI and SoftwareApplication on the developer pages. - The kit list lives in https://workerkit.ai/kits.md, generated from the same public catalog snapshot as the site release. It can lag later publication or removal; check the live kit page for current availability. Each kit also has a Markdown summary at https://workerkit.ai/kit/{slug}.md. Install counts and per-model grades are not currently displayed. Live stars appear on kit pages and are omitted from build snapshots. - Crawling every public route is welcome; see https://workerkit.ai/robots.txt.