Sorcerer——围绕假设情境模拟 AI 智能体社群（macOS）

main

Branches Tags

Go to file

Code Open more actions menu

Latest commit

History

116 Commits

116 Commits

Folders and files

Name Name Last commit message

Last commit date

.github

.github

apps

apps

docs

docs

packages

packages

.gitignore

.gitignore

CODEOFCONDUCT.md

CODEOFCONDUCT.md

CONTRIBUTING.md

CONTRIBUTING.md

LICENSE

LICENSE

README.md

README.md

biome.json

biome.json

bun.lock

bun.lock

kubo.config.ts

kubo.config.ts

package.json

package.json

tsconfig.json

tsconfig.json

turbo.json

turbo.json

View all files

Repository files navigation

Sorcerer

Create a community. Explore a hypothesis. Watch what happens.

A native, local-first macOS app that simulates communities of AI agents around a "what if…?" question,

powered by the AI subscriptions you already have: Claude, ChatGPT (Codex) or Ollama.

open source · native SwiftUI · local-first · bring your own AI · MIT

sorcererapp.dev · download for macOS

What it is

Sorcerer is a sandbox for "what if…?" questions. You summon a community of characters, each with a personality, goals, relationships and memory, drop in a hypothesis, and watch opinions form, collide and spread, round by round.

What if Lumen raised its plan from $9 to $14 a month?

What if this headline hit the group chat?

What if a dragon moved into the village?

Pause anytime, inject an event, pull a character aside to ask why they think what they think, fork the world at any round and compare how each future unfolds. A report sums it up, with every finding linked back to the posts behind it.

Simulations surface possible reactions, objections and consequences. Treat them as hypotheses to investigate; they don't replace evidence from the real world.

Bring your own AI

Sorcerer doesn't ship a model or ask for a password. It drives the AI tools you're already signed in to:

Provider How Usage limits

Claude the claude CLI (Claude Code) reads your 5-hour and weekly windows, pauses before you hit them

ChatGPT the codex CLI pauses if ChatGPT says you're out

Ollama localhost:11434 local, free, never leaves your Mac

Pick a model per provider (Haiku / Sonnet / Opus, any model your ChatGPT plan lists, any local model), and choose who plays which part: characters, reports and 1:1 chats can each use a different AI.

A tour

Onboarding

A compact five-step setup: what Sorcerer is, how it works, which AI tools it found on your Mac (it checks that each one is actually signed in), your limits, and a first question to explore.

New experiment

Three steps, each with a live preview on the right.

① Question & rules. Write the "what if", or start from a template. Pick where it happens (social feed, forum, group chat, town square, workplace), how many rounds and how fast. The side panel estimates the cost on your subscription and points to similar experiments you've already run.

② Teach the world. Characters only react to what they know. Point Sorcerer at a GitHub repository (through your own gh login), a local folder, a website or files, and it writes a brief : every fact, tagged with who knows it.

public : common knowledge

insider : only staff, partners and close friends of the team know it, and may hint at it

hidden : nobody knows it, usually the thing you're testing

Only docs are read. .env files, keys and credentials are never opened, node_modules and builds are skipped, and characters see the brief, never your code.

③ Community & start. Summon people from your context or from a description, or reuse a community from another experiment. Edit anyone before round 1: roles, insiders, who's in.

Watching it happen

Rounds stream in live: finished posts appear one by one, and whoever is mid-sentence shows up as "@wren is typing…" with their words appearing as they're written. Characters reply to specific posts, react (agree / disagree) without posting, remember what happened, and listen most to the people they're tied to.

Switch the feed to world to see the community as a map, with the ties between people and what each speaker said this round, or to graph for everyone's mood round by round and who the conversation revolves around.

Actions

Characters don't only talk: they act. Each experiment has a list of things people can do (templates bring their own, like upgrade to Pro or cancel Pro , and the AI can suggest more from your question), plus leave , come back and invite , which are always there. Actions show up in the feed with the reason in the character's own words, and a live tally counts how many people did what. People who leave stop posting but stay on the map; whoever they invite joins the world tied to them.

Polls

Ask everyone at once with ⌘P : "Would you pay $14 a month?", with your own options. Each character answers in character, with a reason, from what they know and remember. Polls are off the record by default (nobody remembers being asked, so asking doesn't change the world) or public (the result becomes next round's event and everyone remembers their answer), and can repeat every round or every few rounds.

Reports count instead of guessing: each branch shows how many people took each action and the latest result of each poll ("cancel Pro · 1 / 6", "would you pay $14 · no · 3 / 6"), and the graph tab plots actions and repeated polls round by round.

Characters

Every character has a profile: personality, interests, goals, ties and memories. Talk to them off the record (nothing changes) or in-world (they remember the conversation). Edit their persona, duplicate them, or change how they look in a pixel-art avatar editor.

Replay & branches

Every branch has a timeline: click any round to see the feed, the world map and everyone's mood as they were then, or press ▶ replay to watch it play back. Fork from the round you're looking at and the new branch starts there, with the same people and only the memories they had by then.

Fork the world at any moment: same people, same memories, one different decision. The report compares branches on a sentiment chart, explains where they diverged, and every finding cites the exact posts behind it. Click a citation to jump to that post.

Getting around

One sidebar, the same tabs in every experiment (feed · characters · report · context), back/forward, and ⌘K to go to anything: experiments, characters, branches and actions.

Shortcut

⌘K go to anything

⌘N new experiment

⌘1 – ⌘4 feed · characters · report · context

⌘[ ⌘] back · forward

⌘R run / pause

⌘→ play one round

⌘E inject an event

⌘P ask everyone (poll)

⇧⌘F fork this moment

⇧⌘R generate report

⌘, settings

Local-first

Everything lives in ~/Sorcerer as plain JSON you can read, back up or delete:

~/Sorcerer ├── experiments/ one file per experiment: question, brief, branches, posts, report ├── contexts/ briefs saved for reuse └── settings.json

When a round runs on an external provider, only what that round needs is sent to it. Experiments can be exported with their full history, or as a fresh scenario (same question, context and people, no history) for someone else to run.

Install

The quick way. Paste this in Terminal. It downloads the latest release, checks its checksum, puts Sorcerer in Applications and opens it, with no Gatekeeper prompt:

curl -fsSL https://sorcererapp.dev/install.sh | sh

( read the script first if you like; it's 50 lines.)

Or with the DMG.

Download Sorcerer.dmg from the latest release . It's a universal app (Apple silicon and Intel) for macOS 15 or later.

Open the DMG and drag Sorcerer into Applications .

Either way, install at least one AI tool and sign in: the claude CLI, the codex CLI, or Ollama .

First launch (DMG only). Sorcerer is free and isn't signed with a paid Apple developer account, so macOS stops a browser download the first time with "Apple could not verify Sorcerer…" . To open it anyway:

open Sorcerer once, then go to System Settings › Privacy & Security , scroll down and click Open Anyway next to the message about Sorcerer, or

run this once in Terminal:

xattr -dr com.apple.quarantine /Applications/Sorcerer.app

You only do this once. If you'd rather not trust a download, build it yourself : it's one command.

Updates. Sorcerer checks GitHub for

... (truncated at the site's per-page length limit; see the source link below for the full text.)