Town teardown: 1105 tools, narrowed down to 177

Web and native Mac, with a real foothold on the machine itself.

Town is the only agent on this bench with a real foothold on your machine, and the only one where a native Mac app carries the same weight as the website.

A 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.

The narrowing goes further than the main assistant. On the account we examined there were 33 specialised assistants, each with its own instructions and its own allow-list; one of them, a background writer, runs with 17 tools.

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.

The Mac app is native, and that's the whole point

It's Swift with a web view inside, not Electron, and the difference shows up in what it can reach: 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, but it should be an informed one, and Town's marketing isn't where you'll find it spelled out.

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. We never got hold of the source for those, so treat anything we might say about their internals as unsaid.

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: the personality 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.

How long any of it survives, and what happens to a long conversation when it gets compressed, we couldn't pin down.

What we couldn't establish

[TODO 保留本节，上线前逐条复核]

• The full instruction set. The main template runs to 24 sections; the member-facing interface won't return the text.
• Where the personality document gets spliced into the request, and in what order the three request fields are assembled.
• The execution bodies of the code-type routines.
• What compression does to a long conversation, and how long anything is kept.
• Who runs the harness the model conversation lives in.

Pricing

[TODO 数据从 data/pricing.json 读，带 as-of 日期渲染。上线前核一次。]

How Town scores

[TODO 接 /how-scoring-works 的五条轴，等 data/scorecard.json 定稿后填]

• 执行位置
• 工具暴露方式
• 记忆载体
• 渠道入口
• 我们没测到什么

The five axes

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.

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.

Where memory lives

Server-side: a wiki, a memory store, and a people model. The personality is a document you can open and edit, not a line buried in the prompt.

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.

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 four newer routines.

The other three

• PinePhones companies on your behalf about bills, refunds and cancellations.
• InstinctLives in your messaging apps rather than in a new interface.
• MuseMeta's personal agent, with a Linux machine of your own behind it.

Compare all four on the scorecard, read how we score, or see why the existing agent benchmarks don't answer this.