--- title: "Town AI teardown: 1105 tools, narrowed down to 177" description: "Town AI lists 1105 tools and shows its main assistant 177 of them. Its Mac app is native and reaches messages, contacts and the screen." url: https://personalagentbench.com/agents/town/ last_updated: 2026-10-22 --- # Town AI teardown: 1105 tools, narrowed down to 177 Web and native Mac, with the broadest native reach into your Mac of the four. Product **town.com** · Last verified **2026-10-22** · Result **figure withheld, reason on the results page**. [Provenance: differing scored subsets; first-batch records unpublished](https://personalagentbench.com/results/). Town AI is a personal assistant with a website and a native Mac app of equal weight, and of the four products we opened, the one with the broadest native reach into your Mac. The basics, since our teardown says nothing about the company. On 19 October 2026 Town [announced a $108 million Series B](https://dealroom.co/news/163936-town-raises-108m-at-1b-valuation-to-make-ai-assistants-work-as-a-team/) at a $1 billion valuation, co-led by Forerunner Ventures and Menlo Ventures, with Andreessen Horowitz and First Round Capital joining. Before that, [Fortune reported](https://fortune.com/2026/06/03/towns-ai-assistants-andreessen-horowitz-forerunner-55-million/) on 3 June 2026 that Town raised a $55 million Series A led by Andreessen Horowitz, and that it was founded in late 2024 by Jean-Denis Gréze, Plaid's former CTO, and Tony Vincent, formerly a Google director of applied AI. The same piece says you can't sign up without connecting your email and calendar, and that the founder puts work at roughly 70 to 80 percent of use. Each user's assistant is a "Townie" with a name and a look of its own. [Town's homepage](https://www.town.com/), checked on 22 October, lists work tools such as Slack among its integrations, an iPhone app and a Mac app, and says you can message your Townie from Slack, WhatsApp, Telegram or iMessage. The Series B came with new features we have not examined. The announcement says Townies can now ask each other questions and hand tasks to one another, sharing only what their owner would and sending decisions that need a person back to one. A [post by Todd Jackson](https://x.com/tjack/status/2112211241464963147), who says he backed the founders from the start, adds live voice conversation, slide decks, purchases through Stripe Link and passwords through 1Password, and assistants you can CC on an email or add to a group text. All of that postdates the account we took apart, so nothing below covers it, and payments and password access widen what the assistant can touch. Town is not "AI Town", the unrelated virtual town of AI characters that shares its search results. - The client ships a catalogue of 1105 tools across 84 categories. The main assistant is offered roughly 177 of them. - Four separate filters sit between "in the catalogue" and "callable right now", and a tool that fails them returns an empty result rather than a refusal. - On the account we examined there were 33 specialised assistant records. Twenty-nine carry their own instructions; the other four are shells whose work is done in code we never saw. One background writer runs with 17 tools. - The Mac app is Swift with a web view, not Electron, and it reaches messages, contacts, screen and microphone. - The personality is its own document rather than a line buried in the instruction template. For buying context, [Town publishes its plans](https://www.town.com/pricing). An [eesel review](https://www.eesel.ai/blog/town-ai-review) highlights uncertainty about per-action credit costs. Eesel sells its own AI product, so that review is a commercial perspective, separate from our runtime findings. ## Town AI's catalogue of 1105, and you'll never see most of it The client ships a directory of 1105 tools across 84 categories. The main assistant is offered roughly 177 of them. That gap isn't sloppiness; it's the design. Being in the catalogue and being callable in this particular conversation are different states, and four things sit between them: which accounts you've connected, how the agent is configured, what role it's running as, and an explicit allow-list. A search over the catalogue returns schemas only for tools that survived all four. Ask about one that didn't and you get an empty result rather than a refusal, which is a subtle and slightly infuriating distinction if you're trying to work out what the thing can actually do. An empty result is indistinguishable from "no such tool", so the shape of what has been withheld is itself withheld. The narrowing goes further than the main assistant. On the account we examined there were 33 specialised assistant records. Twenty-nine carry their own instructions and allow-lists; the other four are thin shells whose work is done in code we never saw. One of them, a background writer, runs with 17 tools from its allow-list. Their permissions and effects determine its reach; tool count alone does not establish containment. Web search is one of the main assistant's tools, and Town doesn't say who provides it. The tool's own filters hint at an answer: five of the seven result categories it accepts, including the unusual "personal site" and "financial report", are the categories [Exa's search API](https://exa.ai/docs/reference/search) documents, and it filters by published date the same way. That is a match of shape, not a confirmed vendor; we found no Exa name or address on the account we examined. Instinct's search tool returns highlighted excerpts with domain and date filters, which fits the same API less specifically. ## Skills load into this conversation, not a new one A skill in Town is a document, and reading it pulls it into the session you're already in. No fresh context, no handoff, no subprocess. Delegation is a separate mechanism with its own verb, and when work gets delegated the child doesn't inherit your chat wholesale; it gets a task and whatever context was named for it. Keeping those two ideas apart matters more than it sounds. Loading a skill changes what the current conversation knows. Delegating changes who is doing the work. Products that conflate them end up with subagents that have either too much of your history or too little of it. ## The Town AI Mac app is native, and that's the whole point The examined app uses Swift with a web view. Separately, we observed OS access to your messages, your contacts, your screen, your microphone. Web and Mac stay in sync through a shared backend, while the model conversation itself runs somewhere else entirely, over its own channel, in a harness neither client hosts. So the model still isn't running on your laptop. What is on your laptop is a native process with real system access, taking instructions from a model that isn't. Whether that's the right trade is a question for you, and it should be an informed one. Reading your messages and watching your screen is exactly what makes an assistant useful at the things assistants are supposed to be useful at, and it's also the broadest local permission surface of the desktop apps we examined. Town's marketing is not where you'll find that spelled out. ## Town AI routines: the clock can start it A conversation isn't the only way in. A routine can fire on a schedule, on an incoming email, on a calendar entry, on a recording, or because you pressed Run. Most of them are a single unattended model session. Some are a different species entirely: a code type that executes a TypeScript entry point rather than prompting a model. That second kind is worth pausing on. A routine that runs code is not an agent making a judgement call; it's a program, with a program's determinism and a program's blind spots. Mixing both under one word in the interface means "routine" tells you when something runs and nothing about what kind of thing it is. We never got hold of the source for the code type, so treat anything we might say about their internals as unsaid. ## Town AI's memory is server-side, and the personality is a document There's a wiki of assembled background about you, a store of cross-session facts scoped either globally or to a single routine, a library holding documents and routine output, and a people model tracking contacts and the approvals that pass between assistants. The part worth stealing is the personality. It lives in its own document, separate from the assistant's instruction template. Most products bury tone in a line of the system prompt where nobody can find it, let alone change it. Town treats it as a thing you can open, which means "make it less chatty" is an edit rather than a support ticket. How long any of it survives, and what happens to a long conversation when it gets compressed, we couldn't pin down. Compression policy is one of those details that nobody publishes and everybody discovers eventually, usually by noticing the agent has forgotten something it definitely knew last month. Splitting memory across four stores is a real design decision rather than an accident of growth. A wiki is for background that changes slowly. A memory store is for facts that arrive one at a time. A library is for artefacts with a shape of their own. A people model is for the relationships between all of it. Products that use one store for all four jobs end up retrieving a document when they wanted a fact, and the retrieval quality problem that follows is usually blamed on the model. ## What seventeen tools actually buys you The background writer running with 17 tools is the most interesting number on this page, and it's easy to skim past. The writer's allow-list is narrower than the main assistant's. We have not published an analysis of those 17 tools' permissions or effects, so we cannot establish that each is necessary or that the list grants only the reach its job requires. Now consider what happens when that writer misbehaves, whether through a bad instruction, a poisoned document it was asked to summarise, or a plain bug. The tool allow-list narrows the actions offered to that writer. Tool count alone does not establish a small impact: the permissions and effects of each action still matter. That is the reason to examine per-agent allow-lists, and Town applies it to each of the 29 assistant records that carry their own instructions. The main assistant's 177 is the other end of the same idea. It's a lot, and it's a fraction of 1105. Somebody sat down and decided which fraction. ### The catalogue is the marketing, the reachable set is the product There's a trap in how tool counts get quoted, and Town is the clearest example of it on this bench. A product can truthfully say it supports 1105 tools. The catalogue lists 1105 entries; we did not establish that every entry is documented and callable. What that sentence doesn't tell you is that the assistant you're talking to sees roughly one in six of them, that your connected accounts decide part of the rest, and that a tool which fails the filters is indistinguishable from a tool that never existed. None of that is dishonest. It is the same gap that shows up on every product here in a different costume: Pine's 26 names without schemas, Instinct's three defensible totals, Muse's 246 permissions that carry no parameters. A tool count is a statement about a catalogue. What you want is a statement about a session, and no product on this bench publishes one. That's why our scorecard asks how tools are exposed rather than how many there are. The number is the least interesting part of the answer. There's a practical version of this for anyone actually evaluating Town. Don't ask what it integrates with; ask what the assistant you'll be talking to can reach after you've connected the accounts you're willing to connect. Those are different questions with different answers, and only the second one describes your experience of the product. The same test works on every agent you're comparing, and it's the fastest way to cut through a feature grid. ## What we couldn't establish - The full instruction set. The main assistant template we collected has 11 sections. A fuller version said to run to 24 never came back through the member-facing interface, and we have no list of its headings. - Where the personality document gets spliced into the request, and in what order the three request fields are assembled. - The execution bodies of its routines. - What compression does to a long conversation, and how long anything is kept. - Who runs the harness the model conversation lives in. Town's [pricing page](https://www.town.com/pricing) names morning briefings, inbox triage and drafts, meeting briefings, and custom routines. Those are vendor-described uses, separate from our runtime findings. [Instinct vs Town](https://personalagentbench.com/instinct-vs-town/), [Pine vs Town](https://personalagentbench.com/pine-vs-town/) and [Muse vs Town](https://personalagentbench.com/muse-vs-town/) compare its execution and tool access; [family scheduling](https://personalagentbench.com/tasks/family-scheduling/) explains an errand where account reach matters, and [proactive AI agents](https://personalagentbench.com/proactive-ai-agents/) sets its routines beside other products that act unasked. ## FAQ **What is Town AI?** A personal assistant with a website and a native Mac app of equal weight. The model runs server-side in a harness neither client hosts, while the Mac app provides local access to messages, contacts, screen and audio. **How many tools does Town AI have?** The client catalogue lists 1105 tools across 84 categories. The main assistant is offered roughly 177 of them, after four filters: connected accounts, agent configuration, role, and an explicit allow-list. **Is the Town Mac app just a web wrapper?** No. It's Swift with a web view inside rather than Electron, and it reaches system surfaces a browser cannot: your messages, your contacts, your screen and your microphone. **Can Town AI do things without me asking?** Yes. Routines fire on a schedule, an incoming email, a calendar entry, a recording, or a manual run. Most are one unattended model session; some execute code instead of prompting a model. **Where does Town AI store what it knows about me?** Server-side, across a wiki, a cross-session memory store, a library of documents and routine output, and a people model. Unlike Muse, where memory is kept as markdown files on a per-user VM, none of it is stored as files. We could not verify whether users can open Muse's files themselves. **Can I change how Town AI talks to me?** Yes, and more easily than on most products here. The personality is a separate document rather than a line inside the instruction template, so adjusting tone is an edit rather than a request to support. ## Where Town AI lands on the personal agent scorecard Town reaches further into your machine than the other desktop apps we examined, and it carries a lot of machinery for limiting that reach. A catalogue of 1105 tools narrowed by four filters to 177, thirty-three specialised assistant records, most with their own allow-lists, and a background writer trusted with seventeen tools is a product that has clearly thought about blast radius. It also puts a native process on your laptop that can read your messages. Both of those things are true at once, which is why the [scorecard](https://personalagentbench.com/) puts all four products on the same six axes instead of ranking them on a single number. ## Town AI compared with the others we opened Side by side on the same axes, against [Pine](https://personalagentbench.com/pine-vs-town/), [Instinct](https://personalagentbench.com/instinct-vs-town/) and [Muse](https://personalagentbench.com/muse-vs-town/). We only build these for pairs we have taken apart ourselves, because a comparison with marketing on one side is the thing this site exists to replace. ## Errands we describe Town AI against Each errand page says what the job demands and what Town brings to it, next to the other products. They describe; they don't score this product on that errand. - [**Booking a table**](https://personalagentbench.com/tasks/book-a-table/) - [**Getting a subscription cancelled**](https://personalagentbench.com/tasks/cancel-a-subscription/) - [**Disputing a charge**](https://personalagentbench.com/tasks/dispute-a-charge/) - [**Rearranging a family week**](https://personalagentbench.com/tasks/family-scheduling/) - [**Claiming for a delayed flight**](https://personalagentbench.com/tasks/flight-delay-compensation/) - [**Watching something live**](https://personalagentbench.com/tasks/live-alerts/) - [**Organising work files**](https://personalagentbench.com/tasks/work-files/) ## What Town AI costs Five tiers including a real free one. Pro and above are priced per user. Trial, included allowances and annual prices rechecked on 12 October 2026; monthly paid prices remain the 23 September snapshot because the fetched page exposed annual views. | Plan | Price (monthly: 2026-09-23; annual: 2026-10-12) | What it includes | | --- | --- | --- | | Free | $0 · no credit card required | 30 chat messages/month. 7-day Pro trial on signup | | Starter | $15/month · billed monthly on both views | 1,250 credits/month. 10 person research lookups per month | | Pro | $59/month · $49/month per user | 5,000 credits/month. 50 person research lookups per month | | Power | $119/month · $99/month per user | 12,000 credits/month. 150 person research lookups, roughly 2.4x the credits at lower per-credit cost | | Power Plus | $239/month · $199/month per user | 40,000 credits/month. unlimited person research lookups, 2 business day SLA | Monthly paid prices: 23 September 2026 snapshot. Annual prices, trial and allowances: 12 October 2026 check. See the table notes for scope. ## Town AI, axis by axis ## Town AI: where it runs Server-side sync and session handling. The Mac app is native and genuinely reaches into the system: messages, contacts, screen and audio. ## Town AI: how tools are exposed A catalogue of 1105, narrowed in four stages before anything is offered. The main assistant sees 177 of them; one background writer is allowed 17. ## Town AI: where memory lives Server-side: a wiki, a memory store, a library of documents and routine output, and a people model. The personality is a document you can open and edit, not a line buried in the prompt. ## Town AI: what you get to approve Approval is configuration rather than a prompt. The run mode, the account scope, the tool list and the approval mode together decide whether a task may start at all and whether it may write to anything outside, and an individual tool can carry extra approval on top of that. Approval records are kept alongside the people the assistants deal with. Memory has its own add, update and delete tools rather than a setting. ## Town AI: how you reach it Web and Mac carry equal weight, and routines can be started by a clock, an email or a calendar entry rather than by you typing. Town's own site also lists an iPhone app, Slack, texting from any phone, WhatsApp and Telegram, and its help pages say the three chat channels are for asking, since Town cannot reliably message you first there; we did not examine those. ## Town AI: what we could not establish The full instruction set (the member-facing interface refuses it), where the personality document is spliced in, and the execution bodies of its routines. ## The other teardowns - [**Pine**Phones companies on your behalf about bills, refunds and cancellations.](https://personalagentbench.com/agents/pine/) - [**Instinct**Lives in your messaging apps rather than in a new interface.](https://personalagentbench.com/agents/instinct/) - [**Muse**Meta's personal agent, with a Linux machine of your own behind it.](https://personalagentbench.com/agents/muse/) Every runtime we have opened, side by side on the same axes. [See the scorecard](https://personalagentbench.com/) Or read [how we score](https://personalagentbench.com/how-scoring-works/), and [why the existing agent benchmarks](https://personalagentbench.com/benchmarks/) don't answer this. --- Published by Personal Agent Bench (https://personalagentbench.com), run by the Zentor team. Zentor builds agent software of its own, so its products are not on the scorecard and will not be added.