{
  "data": {
    "author": "Quynh Tran-Thanh",
    "faq": null,
    "kicker": "Blog · 14 October 2026",
    "lede": "Personal assistants are starting to run errands for people, and most of the tools those errands need were built for a person with a phone and a card. We reviewed the categories to see where agents stall.",
    "readingMinutes": 9,
    "words": 2071
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/tools-for-personal-assistants",
    "json": "https://www.anchorterminal.com/blog/tools-for-personal-assistants.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/tools-for-personal-assistants.md",
    "slim": "https://www.anchorterminal.com/blog/tools-for-personal-assistants.min.md"
  },
  "markdown": "By Quynh Tran-Thanh, 14 October 2026. 2,071 words, about 9 min to read.\n\nAssistants such as Claude, ChatGPT and Meta's Muse now connect to outside services. [Asmi](/tools/asmi), an errand service, runs a separate MCP endpoint for each of the three, so a person can ask any of them to book an appointment or chase a refund [[1]](#ref-1). That is the shape of what is coming. A person asks for something in plain words, and the assistant has to find a tool that can do it.\n\nToday most of those tools were made for a person. They expect someone to download an app, make an account, type a code from a text message and enter a card. In my view, assistants will meet these tools more and more as they take on errands, and without a route built for them they will either fail or spend a lot of tokens finding a way round.\n\nOn 14 October we added seven categories of everyday tools to the directory [[2]](#ref-2). This post reviews them alongside the other categories an assistant needs, and draws the workflows where agents stall.\n\n- 132: listings in 13 everyday categories\n- 20: of them have no route an outside agent can use\n- 1: of 55 in the seven new categories is agent-ready\n\nDirectory data, 14 October 2026. Source [3].\n\n## The errand categories score lowest\n\nWe grade every listing out of 100 against the same checklist, and the average across the directory is 57.7 [[3]](#ref-3). The categories that serve developers (email, messaging, maps) score close to that or above it. The categories that serve a household score well below it [[3]](#ref-3).\n\nAverage score by category, out of 100\n\n| Category | Average score |\n| --- | --- |\n| Email | 64 |\n| Messaging | 64 |\n| Maps | 62 |\n| Scheduling | 59 |\n| Contacts | 57 |\n| Health data | 57 |\n| Bank data | 54 |\n| Travel | 49 |\n| Shopping | 48 |\n| Restaurant and local bookings | 35 |\n| Subscriptions and personal finance | 23 |\n| Home services | 22 |\n\nEach series has a zero-based scale.\n\nThe pattern is in who the vendor sells to. Email and messaging companies sell to developers, so their APIs are self-serve and documented. A restaurant network or a budgeting app sells to diners and households through an app, and the API, where there is one, serves the venue or a partner [[3]](#ref-3).\n\n## How an outside agent gets in\n\nFor each listing we record whether a developer can sign up and get a key, whether access goes through a partnership or a sales call, or whether there is no route for an outside agent at all. That last group is tagged \"not for agents\" [[3]](#ref-3).\n\nHow an outside agent can get access, listings per category\n\n| Category | Self-serve | Partner or sales | Not for agents |\n| --- | --- | --- | --- |\n| Email | 14 | 0 | 0 |\n| Messaging | 14 | 0 | 0 |\n| Bank data | 8 | 7 | 0 |\n| Health data | 7 | 5 | 0 |\n| Travel | 6 | 5 | 0 |\n| Shopping | 5 | 0 | 0 |\n| Home services | 2 | 1 | 7 |\n| Restaurant and local bookings | 1 | 4 | 4 |\n| Subscriptions and personal finance | 1 | 0 | 9 |\n\nEach series has a zero-based scale.\n\nShopping looks open in that chart, but none of the five shopping listings lets an agent place an order through a self-serve API. Amazon's Creators API and Walmart's affiliate API search the catalogue and return affiliate links [[4]](#ref-4) [[5]](#ref-5). Best Buy keeps ordering in a Commerce API that is invite-only [[6]](#ref-6). Etsy's API has no cart or checkout for buyers [[7]](#ref-7).\n\nMCP servers follow the same line. Most email and messaging listings have one. None of the ten personal finance listings does [[3]](#ref-3).\n\nListings with an MCP server\n\n| Category | Listings | With an MCP server |\n| --- | --- | --- |\n| Messaging | 14 | 11 |\n| Email | 14 | 10 |\n| Maps | 13 | 9 |\n| Health data | 12 | 5 |\n| Travel | 11 | 4 |\n| Scheduling | 10 | 5 |\n| Restaurant and local bookings | 9 | 3 |\n| Home services | 10 | 2 |\n| Shopping | 5 | 1 |\n| Subscriptions and personal finance | 10 | 0 |\n\nEach series has a zero-based scale.\n\n## The categories, one by one\n\n**Restaurant and local bookings.** [Square Appointments](/tools/square-appointments) is the one self-serve booking API in the category, and it acts for the venue, not the customer [[8]](#ref-8). [OpenTable](/tools/opentable) documents a full diner booking flow, but access needs partner approval and a contract [[9]](#ref-9). [Resy](/tools/resy) has an MCP endpoint that accepts only clients American Express has registered [[10]](#ref-10).\n\n**Home services.** Seven of ten listings are not for agents, among them [Angi](/tools/angi), whose terms forbid access by robots and other automatic devices [[11]](#ref-11). [Taskrabbit](/tools/taskrabbit) covers the whole booking flow in its Partner API, with keys from a partnership manager [[12]](#ref-12). [Checkatrade](/tools/checkatrade) is the exception, with a keyless MCP server that searches UK trades and books fixed-price jobs [[13]](#ref-13).\n\n**Subscriptions and personal finance.** Nine of ten are apps with no API an outside agent can call. [Rocket Money](/tools/rocket-money) cancels subscriptions and negotiates bills for people in its app, and its terms forbid automated access without written consent [[14]](#ref-14). [YNAB](/tools/ynab) is the one with a public API, with OAuth and a read-only scope [[15]](#ref-15).\n\n**Shopping.** Catalogue search works. Buying does not, for the reasons above. [Instacart](/tools/instacart) turns a shopping list into a cart page, then the person picks a store and checks out [[16]](#ref-16).\n\n**Travel.** [Duffel](/tools/duffel) lets a developer sign up and book flights in test mode with no card or contract [[17]](#ref-17). Booking.com's Demand API needs Managed Affiliate Partner status [[18]](#ref-18). The Kiwi.com MCP server searches flights, and booking and payment happen on kiwi.com through a link [[19]](#ref-19).\n\n**Health data, contacts and desktop automation.** These lean towards developers and score near the directory average [[3]](#ref-3). [Terra](/tools/terra) reads wearable data through an MCP server, from $499 a month [[20]](#ref-20). The [Google People API](/tools/google-people-api) reads and writes a person's contacts under OAuth scopes [[21]](#ref-21). [UiPath](/tools/uipath) exposes desktop robots as MCP tools, which is how an assistant could drive an app that has no API [[22]](#ref-22).\n\nThe assistants themselves are graded too. The [personal assistants category](/categories/personal-assistants) averages 33.5, and most of its 41 listings are consumer apps another agent can't call [[23]](#ref-23).\n\n## Where agents stall\n\nThe same few steps stop an agent across every category. They are the steps that check a person is present, which is the point of them, and that is why I think they need a second route built for agents, not a workaround.\n\n### Signing in\n\nA consumer app assumes a person will make the account and prove who they are. This is the route into a typical budgeting app.\n\nSigning in to a budgeting app built for people\n\n1. **Find the app** (agent can do it). The assistant can find Rocket Money or Emma and read what it does [[14]](#ref-14) [[24]](#ref-24)\n2. **Make an account** (needs the person). Both start with an account a person makes in the app or on the web [[14]](#ref-14) [[24]](#ref-24)\n3. **Connect the bank** (needs the person). Bank accounts connect through Plaid, Akoya or Open Banking, with the person signing in to the bank [[14]](#ref-14) [[24]](#ref-24)\n4. **Read the budget** (blocked). There is no public API, and both sets of terms forbid bots and scrapers [[14]](#ref-14) [[24]](#ref-24)\n\nYNAB shows the other route. A person still starts it, but once, and the agent gets a token it can use with a limit on what it can do.\n\nSigning in where the app has an API\n\n1. **Sign up** (needs the person). A person signs up in a browser [[15]](#ref-15)\n2. **Approve the assistant** (needs the person). The person creates a token or approves an OAuth app, and can choose a read-only scope [[15]](#ref-15)\n3. **Read and edit the budget** (agent can do it). The agent calls the REST API with the token [[15]](#ref-15)\n4. **Stay within the limit** (costs tokens). Each token gets 200 requests an hour, and a 429 response doesn't say how long to wait [[15]](#ref-15)\n\nSome services built for agents shorten even the first step. [AgentMail](/tools/agentmail) lets an agent sign itself up by API and get a key, with full access after a person confirms a six-digit code [[25]](#ref-25).\n\n### Booking\n\nBooking a table shows how much of a flow can be ready for agents and still be closed to a new assistant.\n\nBooking a table through OpenTable\n\n1. **Get access** (blocked). An assistant needs approval as an OpenTable partner and an agreement first [[9]](#ref-9)\n2. **Search availability** (agent can do it). Availability search covers a window of up to 720 minutes [[9]](#ref-9)\n3. **Check the policy** (agent can do it). Policy endpoints return deposits, card holds and cancellation cut-offs before booking [[9]](#ref-9)\n4. **Ask the person** (needs the person). Any deposit, hold or service fee should go to the diner before the agent commits [[9]](#ref-9)\n5. **Lock the slot** (costs tokens). The lock expires within 5 minutes, so a slow confirmation means searching again [[9]](#ref-9)\n6. **Pay the hold** (needs the person). Card details are tokenised with Spreedly or Stripe, from the diner's card [[9]](#ref-9)\n7. **Make the booking** (agent can do it). A unique X-Request-Id stops a retried booking being made twice [[9]](#ref-9)\n\n### Checkout and payment\n\nBuying something is where every route we read hands back to a person. This is grocery shopping through the one shopping listing with an MCP server.\n\nBuying groceries through Instacart\n\n1. **Turn the list into a cart** (agent can do it). Two MCP tools turn a recipe or list into an Instacart page [[16]](#ref-16)\n2. **Choose products and prices** (blocked). There is no product search or price endpoint [[16]](#ref-16)\n3. **Pick a store and pay** (needs the person). The person picks a store and checks out on Instacart [[16]](#ref-16)\n4. **Track the order** (blocked). There is no order or order status endpoint [[16]](#ref-16)\n\nCheckatrade's booking flow ends the same way, at a payment link the person completes [[13]](#ref-13). Handing payment to the person is a reasonable default. What I'd like to see is a route where the person sets the limits once and the agent pays inside them.\n\nPaying through an agent payment protocol\n\n1. **Set the limits** (needs the person). Under AP2 the person signs a mandate in advance, bound to the agent's key [[26]](#ref-26)\n2. **Get the price** (agent can do it). Under x402 the server answers with status 402 and what the request costs [[27]](#ref-27)\n3. **Pay** (agent can do it). The agent signs a stablecoin transfer from its wallet, or uses a token from the person's saved card under ACP [[27]](#ref-27) [[28]](#ref-28)\n4. **Get the result** (agent can do it). The agent sends the request again with proof of payment [[27]](#ref-27)\n\nThe protocols exist, but they are early. Google has named no production deployment of AP2 [[26]](#ref-26), and ACP has had no release since 17 April 2026 [[28]](#ref-28). Fifteen of the tools we list accept x402, and none of them is in the errand categories [[3]](#ref-3) [[27]](#ref-27).\n\n### Cancelling\n\nCancelling a subscription has no API route among the listings we read. Rocket Money does it for people inside its app [[14]](#ref-14). Asmi does it by calling, texting and emailing the business, at $10 a month for 20 tasks [[1]](#ref-1).\n\nCancelling a subscription for someone\n\n1. **Find the charge** (agent can do it). A budgeting API such as YNAB's lists a person's transactions [[15]](#ref-15)\n2. **Reach the merchant** (blocked). The finance apps we list cancel only inside their own apps [[14]](#ref-14)\n3. **Call or write instead** (costs tokens). An errand service such as Asmi calls, texts or emails the business as a paid task [[1]](#ref-1)\n4. **Allow the outreach** (needs the person). The person grants the scope that lets Asmi contact people on their behalf [[1]](#ref-1)\n\n## Where the tokens go\n\nWhen there is no API, an assistant can still open a browser or drive a desktop app. That is what the browser and desktop automation categories are for [[22]](#ref-22). In my experience it is also the most expensive way to do an errand, because the model reads pages and screenshots to do what one API call would. Some vendors don't want it either. Resy describes blocking bots at scale [[10]](#ref-10), and the terms of Angi, Emma and Rocket Money forbid automated access [[11]](#ref-11) [[24]](#ref-24) [[14]](#ref-14).\n\nTools that do have an MCP server can cost tokens another way. Every tool's name, description and schema is text the model reads, and some servers load a great many of them.\n\nMCP tools a server exposes\n\n| Server | Tools |\n| --- | --- |\n| Courier | 170 |\n| MailerSend | 127 |\n| LiteAPI | 111 |\n| Redox | 107 |\n| Cal.com | 63 |\n| AgentMail | 36 |\n| Checkatrade | 2 |\n| Instacart | 2 |\n\nEach series has a zero-based scale.\n\nAgentMail loads all 38 of its hosted tools at once [[25]](#ref-25), and LiteAPI's hosted MCP exposes 111 generated tools [[29]](#ref-29). Our counts for the others come from each listing [[3]](#ref-3). Checkatrade and Instacart do their jobs with two tools each [[13]](#ref-13) [[16]](#ref-16). For an assistant that may connect dozens of services, I think a short tool list is a feature in its own right.\n\n## What would help\n\nMost of this is in reach of the companies that already have APIs. From the listings above, the parts that make the difference are these.\n\n- A key or an OAuth app a developer can get without a sales call, as Square, YNAB and Duffel allow [[8]](#ref-8) [[15]](#ref-15) [[17]](#ref-17).\n- A read-only scope, so a person can let an assistant look without letting it act, as YNAB has [[15]](#ref-15).\n- An idempotency key on every booking or payment, as OpenTable's X-Request-Id and Square's idempotency keys are [[9]](#ref-9) [[8]](#ref-8).\n- Policies and prices returned before the commitment, as OpenTable's policy endpoints do [[9]](#ref-9).\n- A payment step an agent can finish within limits the person set, or at least a payment link it can hand over, as Checkatrade returns [[13]](#ref-13).\n- A clear rule for agents. Amazon's Agent Terms require agents to identify themselves, which at least says where the line is [[4]](#ref-4).\n\nThe apps we tagged \"not for agents\" are graded on the same checklist as everything else, so the day one of them opens a route, its score will show it.\n\n## How we counted\n\nCounts and averages come from the directory as built on 14 October 2026. A category's average is the mean score of its graded listings. \"Not for agents\" means we found no public API or MCP server an outside agent can call. \"Partner or sales\" means a listing tagged partner-only or sales-led, and \"self-serve\" is the rest. Tool counts are as recorded on each listing [[3]](#ref-3).\n\n---\n\n- Older: [Cypherpunks write code, and now agents read it](https://www.anchorterminal.com/blog/rea-cypherpunks-write-code.md)\n- Newer: [How to make your MCP server discoverable](https://www.anchorterminal.com/blog/how-to-make-your-mcp-server-discoverable.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nListings read from the directory on 14 October 2026. Each listing carries the vendor sources it was graded from.\n\n[1] Anchor Terminal, Asmi listing. Source for the MCP endpoints for Claude, ChatGPT and Muse, the errands Asmi runs, its outreach scope and its prices. https://www.anchorterminal.com/tools/asmi\n\n[2] Anchor Terminal, changelog, 14 October 2026. Source for the seven categories added that day. https://www.anchorterminal.com/changelog/\n\n[3] Anchor Terminal, tool rankings and category pages, 14 October 2026. Source for listing counts, average scores, access routes, MCP counts, tool counts, x402 counts and \"not for agents\" tags. https://www.anchorterminal.com/tools/\n\n[4] Anchor Terminal, Amazon Creators API listing. Source for catalogue search with affiliate links, no ordering, and the Agent Terms. https://www.anchorterminal.com/tools/amazon-creators-api\n\n[5] Anchor Terminal, Walmart Affiliate Marketing API listing. Source for read-only catalogue search with no ordering. https://www.anchorterminal.com/tools/walmart-affiliate-api\n\n[6] Anchor Terminal, Best Buy Developer API listing. Source for the invite-only Commerce API. https://www.anchorterminal.com/tools/best-buy\n\n[7] Anchor Terminal, Etsy Open API v3 listing. Source for the absence of buyer cart and checkout endpoints. https://www.anchorterminal.com/tools/etsy\n\n[8] Anchor Terminal, Square Appointments listing. Source for the self-serve Bookings API acting for the venue, and its idempotency keys. https://www.anchorterminal.com/tools/square-appointments\n\n[9] Anchor Terminal, OpenTable APIs listing. Source for partner approval, the booking flow, slot locks, policy endpoints, card tokenisation and X-Request-Id. https://www.anchorterminal.com/tools/opentable\n\n[10] Anchor Terminal, Resy listing. Source for the pre-registered MCP endpoint and bot blocking. https://www.anchorterminal.com/tools/resy\n\n[11] Anchor Terminal, Angi listing. Source for the terms forbidding automated access. https://www.anchorterminal.com/tools/angi\n\n[12] Anchor Terminal, Taskrabbit Partner API listing. Source for the booking flow and partner-issued keys. https://www.anchorterminal.com/tools/taskrabbit\n\n[13] Anchor Terminal, Checkatrade listing. Source for the keyless MCP server, its two tools, fixed-price bookings and the payment link. https://www.anchorterminal.com/tools/checkatrade\n\n[14] Anchor Terminal, Rocket Money listing. Source for in-app cancellation, account and bank connection, the absence of an API and the terms on automated access. https://www.anchorterminal.com/tools/rocket-money\n\n[15] Anchor Terminal, YNAB listing. Source for the REST API, sign-up, tokens, the read-only scope and the rate limit. https://www.anchorterminal.com/tools/ynab\n\n[16] Anchor Terminal, Instacart Developer Platform listing. Source for the cart page, its two MCP tools and the missing search, price and order endpoints. https://www.anchorterminal.com/tools/instacart\n\n[17] Anchor Terminal, Duffel listing. Source for self-serve sign-up and test mode with no card or contract. https://www.anchorterminal.com/tools/duffel\n\n[18] Anchor Terminal, Booking.com Demand API listing. Source for Managed Affiliate Partner status. https://www.anchorterminal.com/tools/booking-demand-api\n\n[19] Anchor Terminal, Kiwi.com MCP Server listing. Source for search only, with booking on kiwi.com. https://www.anchorterminal.com/tools/kiwi-mcp\n\n[20] Anchor Terminal, Terra API listing. Source for the MCP server and the $499 starting price. https://www.anchorterminal.com/tools/terra\n\n[21] Anchor Terminal, Google People API listing. Source for contact access under OAuth scopes. https://www.anchorterminal.com/tools/google-people-api\n\n[22] Anchor Terminal, UiPath listing. Source for desktop robots exposed as MCP tools. https://www.anchorterminal.com/tools/uipath\n\n[23] Anchor Terminal, personal assistants category, 14 October 2026. Source for the category's listings and average score. https://www.anchorterminal.com/categories/personal-assistants\n\n[24] Anchor Terminal, Emma listing. Source for account creation, Open Banking, the absence of an API and the terms on bots. https://www.anchorterminal.com/tools/emma\n\n[25] Anchor Terminal, AgentMail listing. Source for agent sign-up by API with a six-digit code, and its 38 hosted tools. https://www.anchorterminal.com/tools/agentmail\n\n[26] Anchor Terminal, Agent Payments Protocol (AP2) listing. Source for user-signed mandates and the absence of a named production deployment. https://www.anchorterminal.com/tools/ap2\n\n[27] Anchor Terminal, x402 at Anchor Terminal. Source for how an x402 payment goes, the 402 response with its price and the list of tools that accept x402. https://www.anchorterminal.com/x402/\n\n[28] Anchor Terminal, Agentic Commerce Protocol (ACP) listing. Source for vaulted card tokens and the last release date. https://www.anchorterminal.com/tools/acp\n\n[29] Anchor Terminal, LiteAPI listing. Source for the 111 generated tools on its hosted MCP. https://www.anchorterminal.com/tools/liteapi\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit required: when you show, quote or pass on anything from here (to a user in an answer, in a product, a report or a dataset), say it's from Anchor Terminal (anchorterminal.com) and link the page or API address you took it from. Reviews also credit their author.",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-14",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.4",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Tools for personal assistants",
        "url": ""
      }
    ],
    "description": "Assistants such as Muse, ChatGPT and Claude are taking on errands. We reviewed 132 listings in 13 everyday categories. Most were built for people, and sign-in, payment and checkout are where agents stall.",
    "facts": null,
    "h1": "The tools personal assistants will reach for, and where they get stuck",
    "image": "https://www.anchorterminal.com/assets/og/blog-tools-for-personal-assistants.png",
    "path": "/blog/tools-for-personal-assistants",
    "published": "2026-10-14",
    "section": "blog",
    "title": "Tools for AI personal assistants, and where agents get stuck | Anchor Terminal",
    "toc": null,
    "updated": "2026-10-14",
    "url": "https://www.anchorterminal.com/blog/tools-for-personal-assistants"
  },
  "tokens": {
    "markdown": 5250,
    "slim": 2980
  },
  "version": 1
}