Personal Agent Bench

Pine AI vs Town AI, a borrowed browser or a Mac foothold

Pine logoPine

Borrows a
browser.

When browser work is needed it is handed to a separate agent on your own machine, and the cloud alternative leases you a remote desktop that you watch over VNC.

26 tool names exposed to the client

Town logoTown

A foothold on
your Mac.

Its Mac application is native and genuinely reaches into the system. Messages, contacts, screen and audio are within range.

177 shown from 1105 catalogued tools

Neither ran the model on your computer in the routes we examined; both send your requests to a server. Where Pine and Town actually run.

Last verified 2026-10-23

Four differences

What it is

Pine Phones companies on your behalf about bills, refunds and cancellations.

Town Web and native Mac, with the broadest native reach into your Mac of the four.

Phone calls

Pine is the one built around it.

Town was not found doing this in our teardown.

When you approve

Pine asks you while the task runs: a start confirmation, and an identity confirmation when the company on the phone wants to hear from the account holder.

Town has you settle it in settings before a task ever starts; an individual tool can carry extra approval on top.

What it remembers

Pine gives you no file to edit. There is a server-side persona and a structured fact layer queried on demand.

Town keeps a wiki, a memory store and a people model on the server, and its personality is a document you can open and edit.

Who starts the work

Pine answers to four request shapes, including a REST endpoint that other people's agents can call.

Town routines can be started by a clock, an email or a calendar entry rather than by you typing.

What the same tasks showed

Two kinds of evidence follow: what we observed in the products, and what happened on shared tasks using adapted runtimes. The task results do not tell us which product decision caused an outcome.

Separate batches of adapted runtimes using a substituted model, not end-to-end results for the consumer apps. No combined score or ranking.

October

20 shared tasks, one attempt each

  1. PineNo data, exam fault
  2. Town15% (5–36) · 20 tasks

September, same tasks

114 shared tasks, older rules and builds

  1. Pine30% (23–36) · 114 tasks
  2. Town37% (30–44) · 114 tasks

September, own draw

Each product answered its own draw, so rates are not directly comparable.

  1. Pine20% (14–25) · 167 tasks
  2. TownFigure withheld

95% interval measured rateProduct runtimes rebuilt for the exam with one substituted model, not the consumer apps measured end to end. Results and caveats

Pine AI vs Town AI is a useful pair to put side by side, because they take different approaches to how far the agent reaches into your own machine. Pine runs its model on its own servers and sends browser work either to a helper on your computer or to a leased remote desktop. Town ships a native Mac application that reaches your messages, your contacts, your screen and your audio. Price, tool exposure and channels are separately observed differences; runtime location does not establish their causes.

  • One of them borrows your browser when it must. The other has a native foothold in your messages, contacts, screen and audio.
  • Neither ran the model on your computer in the routes we examined; both send your requests to a server.
  • Their tool surfaces differ by a factor of forty, and the larger one is deliberately hidden from its own assistant.
  • Town has documented clock and email triggers. Our Pine teardown does not establish whether it can start work without you.
  • Pine has no October completion figure, because its submissions never reached the exam in that batch. Town has one, printed beside the September figures and never merged with them.

The basics first. Pine is sold at 19pine.ai, and its App Store listing names the developer as 19Pine Corp. It publishes its prices: personal and household prices are displayed per week, and work prices per month; these units do not establish renewal cadence. Confirm billing cadence and fees at checkout. Town is made by Town.com, Inc. You sign up on the web, there is a native Mac app, and it publishes its prices, starting with a free plan. Town's own site also lists an iPhone app, Slack, texting from any phone, WhatsApp and Telegram, and says the three chat channels are for asking, since Town cannot reliably message you first there; we have not examined those. Fortune reported on 3 June 2026 that Town raised a $55 million Series A led by Andreessen Horowitz, that it was founded in late 2024 by Plaid's former CTO, and that roughly 70 to 80 percent of its use is for work. The current tiers for both are on our teardown pages, Pine and Town.

The Pine and Town runtimes, side by side

Show the scorecard rows
AxisPineTown
Where it runsServer-side. Neither the web app nor the Mac app runs the model. Browser work is handed to a separate agent on your own machine; the cloud route leases you a remote desktop you watch over VNC.Server-side sync and session handling. The Mac app is native and genuinely reaches into the system: messages, contacts, screen and audio.
How tools are exposed26 tool names reach the client, grouped by job. The schemas never do. Computer use bypasses that list entirely.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.
Where memory livesNo file you can edit. A server-side persona plus a structured fact layer queried on demand.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.
What you get to approveConfirmation is required before a billing action, and there are two separate kinds of it. A start confirmation gates the task and carries its own expiry. A different identity confirmation appears when the company on the phone wants to hear from the account holder: a card naming what to say, which number will be calling, and when the offer lapses. Both the request and the receipt land in the task history. We found no editable local memory file; deletion of server-side facts was not established.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.
How you reach itWeb and Mac, outbound phone calls, a voice copilot, and a REST endpoint other people's agents can call. Four request shapes under one brand.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.
What we could not establishProduction instruction text (roughly 126 probes, all empty), the calling tool's schema, which carrier places the calls, and whether work is split across a fast and a slow model.The full instruction set (the member-facing interface refuses it), where the personality document is spliced in, and the execution bodies of its routines.
Completed, scored subset20%14 to 25 · 167 tasksWithheld
Where it runs
PineServer-side. Neither the web app nor the Mac app runs the model. Browser work is handed to a separate agent on your own machine; the cloud route leases you a remote desktop you watch over VNC.
TownServer-side sync and session handling. The Mac app is native and genuinely reaches into the system: messages, contacts, screen and audio.
How tools are exposed
Pine26 tool names reach the client, grouped by job. The schemas never do. Computer use bypasses that list entirely.
TownA catalogue of 1105, narrowed in four stages before anything is offered. The main assistant sees 177 of them; one background writer is allowed 17.
Where memory lives
PineNo file you can edit. A server-side persona plus a structured fact layer queried on demand.
TownServer-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.
What you get to approve
PineConfirmation is required before a billing action, and there are two separate kinds of it. A start confirmation gates the task and carries its own expiry. A different identity confirmation appears when the company on the phone wants to hear from the account holder: a card naming what to say, which number will be calling, and when the offer lapses. Both the request and the receipt land in the task history. We found no editable local memory file; deletion of server-side facts was not established.
TownApproval 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.
How you reach it
PineWeb and Mac, outbound phone calls, a voice copilot, and a REST endpoint other people's agents can call. Four request shapes under one brand.
TownWeb 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.
What we could not establish
PineProduction instruction text (roughly 126 probes, all empty), the calling tool's schema, which carrier places the calls, and whether work is split across a fast and a slow model.
TownThe full instruction set (the member-facing interface refuses it), where the personality document is spliced in, and the execution bodies of its routines.
Completed, scored subset
Pine20% · 95% interval 14–25% · 167 scored tasks
TownWithheld
The completion row covers differing September scored subsets of adapted product runtimes using the substituted exam model, not shipping consumer products. Underlying first-batch records are unpublished; Muse's rate follows an adapter repair. Result provenance and limitations. Neither figure is a ranking and we never combine them. As of 2026-09-22.

Where Pine and Town actually run

Pine runs server-side. Neither the web app nor the Mac app runs the model. When browser work is needed it is handed to a separate agent on your own machine, and the cloud alternative leases you a remote desktop that you watch over VNC. Two execution places for two kinds of work, inside one product.

Town also handles sync and session work on a server, but its Mac application is native and genuinely reaches into the system. Messages, contacts, screen and audio are within range.

Read those two paragraphs again and notice what they are not saying. Neither product runs its model on your laptop. The difference is not local versus cloud intelligence. It is how much of your machine the cloud intelligence can reach, and there the gap is wide: Pine's local agent drives a browser, while Town's Mac app reaches messages, contacts, screen and audio.

That is why the axis matters rather than being an architecture trivia question.

An agent working in a leased remote desktop does the wrong thing on a machine that is not yours; we did not establish how long that desktop lives. Its local files may be isolated, but authenticated external actions can leave lasting consequences outside the box. An agent with a native foothold doing the wrong thing is doing it to your actual messages and your actual contacts, in a place where the result stays after the task ends.

Neither arrangement is safer in the abstract. They are different exposures, and only one of them is usually described on a product page.

Town's reach buys something real. An agent that can see your screen and read your messages has context no API hands over, and that is the argument for building it that way. The cost is that a bad decision can land on your own computer rather than a leased one.

Pine's Mac route is an important exception to the remote-desktop picture: its browser agent runs on your own machine. It drives a browser, while Town's native reach also covers messages and contacts. A leased desktop is therefore not a complete description of Pine's exposure.

Forty times the tools, most of them hidden

Pine exposes 26 tool names to the client, grouped by job, and never ships their schemas. Computer use bypasses that list completely, which means the published count does not describe what the product can actually do.

Town catalogues 1105 tools and narrows them in four stages before anything is offered. Its main assistant is shown 177. One background writer is allowed 17.

The headline is the ratio, but the interesting part is that both products hide most of their surface, by opposite methods. Pine withholds the schemas, so you can see the names and not the shapes. Town withholds the tools themselves, showing its own assistant a sixth of the catalogue. In both cases the number a comparison article would print, 26 against 1105, is the least informative number available.

Town and Pine on memory: one you can open, and one you cannot

Town keeps a wiki, a memory store and a people model on the server, and its personality is a document you can open and edit rather than a line buried in a prompt. Of the four products we opened, Town is the one where we confirmed a personality document you can edit.

Pine gives you no file to edit. There is a server-side persona and a structured fact layer queried on demand.

The practical difference is correction. When Town has the wrong idea about you, there is a surface where that idea lives and you can go and change it. When Pine has the wrong idea about you, the fact layer is queried on demand and you are asking the product to forget something rather than editing a document.

Pine and Town front doors: examined and advertised channels

Pine answers to four request shapes under one brand: the web and Mac apps, outbound phone calls, a voice copilot, and a REST endpoint that other people's agents can call. That last one deserves a second look, because it means Pine is a component other software can drive, not only a product a person uses.

Town gives web and Mac equal weight, and routines can be started by a clock, an email or a calendar entry rather than by you typing.

Pine's API can be triggered by external software; that does not establish autonomous initiation. Town has documented clock and email triggers, while actions through its local helper depend on that helper being available. Town's clock can wake an agent that reaches your messages. That is the combination to think about, rather than either half alone.

What neither of them will confirm

Both pages on this site carry a section for what we could not establish, and the two lists are instructive.

For Pine we could not get the production instruction text, after roughly 126 probes that all came back empty, nor the calling tool's schema, nor which carrier places the calls, nor whether work is split across a fast and a slow model. For Town we could not get the full instruction set, because the member-facing interface refuses it, nor where the personality document is spliced in, nor the execution bodies of its routines.

The shape of the two gaps is similar. In both cases the thing we could not reach is the instruction text, which is the part that decides behaviour. Other surfaces also remain inaccessible: Pine's carrier and production-model routing, and parts of Town's server-side execution. The instruction text is one evidence gap among several.

Pine vs Town on phone calls

Pine phones companies on your behalf about bills, refunds and cancellations. We did not find Town doing this. Instinct has added a concierge that phones businesses, and Muse has business calling in early access, but Pine is the one built around it.

It is worth separating from the rest of the comparison because it is the only capability here that reaches outside software entirely. Every other difference between these two products is about which machine runs code and what that code can see. A phone call puts an agent into a conversation with a person who did not agree to talk to a machine, about your account.

That changes the failure mode again. A tool call that goes wrong leaves a log. A call that goes wrong leaves a human on the other end with an impression of what you authorised, and the record is a summary written by the system that made the call, unless it went through Pine's developer API, which returns a transcript. We could not establish which carrier places the calls, which is one of the gaps listed above, and for a capability of this kind that gap is larger than it sounds.

Pine's four front doors sit around that call. The REST endpoint means a task can be started by software rather than a person at a keyboard; as above, that is not the same as Pine starting work on its own.

Pine vs Town: which one asks before it acts

Pine asks you while the task runs. Town has you settle it in settings before a task ever starts.

Pine has two separate confirmations. A start confirmation gates the task and carries its own expiry. An identity confirmation appears when the company on the phone wants to hear from the account holder: a card naming what to say, which number will be calling, and when the offer lapses. Both the request and the receipt land in the task history.

Town treats approval as 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. Approval records are kept alongside its model of the people the assistants deal with.

The two fit the jobs they are built for. A call to a company has a person waiting, so Pine's question comes at the moment the company asks for you. A Town routine can be started by a clock or an email when you are not there, so its answer has to be settled in advance. If you use Town, read a routine's settings before it first fires. The control rows for Pine and Town have the detail.

Pine vs Town: which one the choice actually comes down to

Strip out everything that both products claim and two questions remain.

The first is whether you want an agent that can see your computer. If the work you have in mind depends on context that lives on your screen and in your messages, Town's examined Mac app reaches that context, while Pine's helper, as we examined it, drives a browser. If it does not, the foothold is exposure you are carrying for nothing.

The second is whether the errands you care about end in a conversation with a company. Cancellations, refunds and billing disputes often do, and Pine is built around that. Town's distinguishing reach is into your own Mac.

Those two questions are nearly independent, which is why a completion rate is a poor way to choose between them. A rate averages across a task set that mixes both kinds of work. The products are not competing on the same axis, whatever a single number implies, and the parts of each that we could not establish sit in the same place for both.

Put as fit rather than as a ranking:

  • Pine logoPick Pine if the errand ends in a conversation with a company, a cancellation, a refund or a billing dispute, and you want to be asked at the start and again when the company asks for you.
  • Town logoPick Town if the work depends on what is on your Mac, its messages, contacts or screen, and you are willing to decide in advance what its routines may do.

What the Pine and Town numbers do and do not say

Our September scorecard carries a completion rate for Pine, with an interval and a task count. Town's September figure is withheld, because the figure we were given does not reconcile with the records we can check; Muse vs Town explains the five refused batches and roughly 13% across 30 scorable tasks. Pine's figure is not a ranking either, and this page is not an argument that one of them is better. Why we never print a combined score is in how scoring works.

A completion rate measures whether an errand finished. It cannot tell the difference between an errand that finished inside a leased remote desktop and one that finished through a native application holding your messages open. Those are different products to live with, and the rate is blind to the distinction this entire page is about. Read the rate for what it measures and read where it runs for the rest.

On the separate 114-task batch, measured on 22 September 2026, Pine completed 30% (95% interval 23 to 36%; 546 attempts) and Town completed 37% (95% interval 30 to 44%; 513 attempts). Both answered the same 114 tasks. These are replicas of older runtimes with the same GLM 5.3 flash model substituted, under older rules, filtered from roughly 5,600 attempts using eight rules, not fresh product results. Excluding attempts that never reached the exam removes over 400 each for Pine and Muse; if those were refusals rather than broken entry, the supplied alternative estimates are about 22% for Pine and 24% for Muse. Transcripts were supplied but have not been republished here; some files were overwritten or shared between attempts, so the underlying record is incomplete. Keep this measurement separate from the main table and do not treat interval overlap or a point estimate as a significance test.

The October batch, run on 11 October 2026, gave every product the same 20 tasks once each, under the exam rules that took effect at the end of September. Town completed 3 of 20 scored attempts, 15% (95% interval 5 to 36%). Pine has no October figure: almost none of its submissions reached the exam, so its column was set aside rather than scored as zero, and the cause is not yet known and may be on the exam side. The tasks, the rules and our replicas all changed since September, so Town's October figure is a new measurement, not Town getting worse, and the batches are never added together. The other October batches, each with its own tasks and builds, are listed separately on the results page, and individual recorded attempts are on the run pages, as counts rather than rates.

For price context, the Town snapshot includes a free plan, Starter at $15/month, and Pro at $49/month billed annually, checked 2026-11-05. Other monthly paid prices retain their earlier snapshot date, as the pricing note explains.

The Pine pricing snapshot starts personal use at $9.00/week, dated 2026-09-23. Weekly personal pricing is not directly comparable to a monthly work plan; the page separates billing periods and additional fees.

FAQ

What is the main difference between Pine and Town? Where the work happens. Pine runs its model on its servers and, when it needs a browser, hands the work to a helper on your computer or leases a remote desktop. Town's Mac application is native and reaches messages, contacts, screen and audio.

Does either of them run the model on my machine? No. Both think on a server. The difference is how much of your machine that server-side intelligence can reach, not where the intelligence lives.

Which has more tools? Town catalogues 1105 and shows its main assistant 177. Pine exposes 26 names and no schemas, and its computer use bypasses that list anyway. Neither count describes what the product can do.

Can I edit what either one remembers about me? Town's personality is a document you can open and edit, alongside a server-side wiki, memory store and people model. Pine gives you no file: a server-side persona and a fact layer queried on demand.

Can either one act while I am not there? Town can: its routines can be started by a clock, an email or a calendar entry. Pine can be driven by other software through its REST endpoint, and it places outbound calls, but our teardown did not establish whether it starts work on its own.

Who makes Pine and Town, and what do they cost? Pine's terms name 19Pine Pte. Ltd.; the App Store seller is 19Pine Corp. It publishes weekly personal plans and monthly work plans. Town is made by Town.com, Inc. and publishes monthly plans, starting with a free one. Details are on our Pine and Town pages.

Did you test these or read about them? We took both apart at the runtime level. Runtime findings come from the Pine and Town teardowns; completion figures come from the separate exam data on the scorecard.

Other comparisons

All six pairs, and why there are only six, are on the comparisons page.