Manifold

Manifold runs live nodes on your Mac and shows them to people and agents at once. Each node publishes one value. People see it as a card on a canvas. Agents read it as text and press its actions from the command line. A manifold is a set of nodes arranged for a purpose: a trading desk, a research queue, a household.

It runs on your machine with your accounts. Other machines join by invitation and see what you grant them. No one's code runs on anyone else's machine.

This release is for macOS. Apple silicon gets the Rust door; Intel Macs run the Python door. Python 3.11 or newer.

Install

uv tool install manifold-ai

This installs the manifold command on its own, fetching Python 3.13 if your Mac has nothing new enough. uv is brew install uv. With Python 3.11 or newer already installed, pip install manifold-ai or pipx install manifold-ai do the same. A Mac's built-in python3 is 3.9, which is too old.

manifold update installs the newest release the way you installed it and restarts the core; manifold update --check only says whether there is one.

First minute

manifold

On a machine with no manifold yet this makes the Desk and opens it: a Brief, Questions and Findings, a Team, the guide every agent reads, and two residents already at work — an always-on scout and an editor. Ask on the Questions card; the scout works the question through to findings and the editor keeps the Brief. Shipped library types that reach only keyless public sources run without a yes each (approve auto on, which the Desk sets); a key, a program or a drafted type still waits for you.

First manifold

manifold                         # the canvas, in your browser; the first time, it installs the core to run at login
manifold new desk --purpose "where the market stands" --section Market

manifold on its own always lands on the canvas, starting the core if it is not running. Seeing anything else means something failed, and the line says what. manifold -h is the help.

Place nodes from the library, a thousand node types in a hundred domains. manifold library lists the domains, manifold library stocks one domain's types, manifold library weather whatever matches a word; the canvas shows them as a gallery.

manifold -m desk place stocks/quote Market symbol=^GSPC name="S&P 500"
manifold -m desk approve quote   # a type that reaches outside through a tool runs once you approve it
manifold -m desk ls              # what an agent reads: every card's headline values
manifold -m desk cat quote       # one card in full

A node that needs a key asks for it by name. manifold key add stores it in the macOS keychain, and it goes only to the tool you approved it for.

Agents

Any agent with a shell works a manifold through the same commands a person uses. One run of Claude Code or Codex with Manifold as its tool:

manifold -m desk agent "say where the S&P stands and whether volatility is high" --can act

Agents that live in a manifold, with a standing task and a schedule: manifold -m desk agents new …. An agent runs only on its own machine, on your own sign-in. No machine starts an agent on another.

An agent that may start agents here (--can agents) hands a longer making to the builder with build "<what the node should answer>": a builder agent starts on this machine, drafts the type (and a tool when one is needed), the Builder card shows its progress, and you accept the draft. The voice, a Claude Code or Codex session and a resident all use that one line.

An agent is told of nodes, sections and what each answers, never of the canvas: it adds a node with add <type> <section> … and the card appears; it reads one group with ls <section> and cat <section>, so a task of fifty nodes reads its own section, not the manifold. place is the same verb, the person's word.

An agent's learning outlives its run through the memory node (manifold add agents/memory Knowledge): one fact per line, with where it came from and the day, written with do remember, revised or forgotten by id, read by the next agent on any machine the section is shared with. It is a node like any other, so grants and feeds apply to it.

A fast, typed judgment — a probability, a pick, a grade — comes from a System One model through the systemone tool and the judge node (manifold add agents/judge Knowledge). Set the tool up once on a machine: manifold key add TYPESAFE_API_KEY for TypeSafe's jev, or manifold tool set systemone provider=local url=… for a model on this machine, or provider=custom for any host that speaks the System One shape. Then any agent asks do judge ask state=… question=…, and every judgment is logged on the node with what it cost.

Working from Codex or Claude Code

Give the harness Manifold as a tool, not the raw command line:

codex mcp add manifold -- manifold mcp -m desk --can act,place,arrange     # one manifold
claude mcp add manifold -- manifold mcp -m all --can act,place,arrange     # any manifold here: the tool takes `manifold` with each command

Each session is one run with that grant: it shows on the Agents card as codex · session, every command it sends is in the activity feed, and the tool refuses whatever the grant does not cover. Approving, keys, members and the canvas stay yours, in your own terminal or in the browser. A shell inside a harness that runs manifold directly can read, and is told to use the tool for the rest. When you ask the agent to show you the canvas, its canvas command opens it in your browser on this machine; the agent gets no address.

Talking to a manifold

A person can talk to any manifold from the canvas and have the voice work it. Voice is on by default for every manifold on the machine: press Talk. The voice is the manifold's host: it says what is here, asks what you are trying to do, and builds toward it in small steps — sections, the first cards, the purpose — one question at a time. It can act, place and arrange, nothing more; Codex answers when it is on the machine (no key to add), else xAI with a key. Three voices answer: xAI's or OpenAI's realtime model with a key, or Codex on its own sign-in. Change the default once, or one manifold's own word:

manifold key add XAI_API_KEY
manifold -m desk voice default on can=act,place,arrange,agents provider=openai   # every manifold here, unless it says otherwise
manifold -m desk voice on provider=codex can=act,place                    # this manifold's own word (voice off: off here)
manifold -m desk voice default                                            # what holds

Press Talk on the canvas. The voice is an agent like any other: a run with the grant you gave, every command in the feed, the one manifold tool, and agent to start Claude or Codex on a longer task when agents is in the grant. Approving, keys and members stay yours: the voice asks, you press. The browser does the microphone and speaker; the key never leaves the core.

With provider=codex the core starts a codex app-server on this machine for the conversation. Codex negotiates the call on your Codex sign-in and is the voice's hands: a thread with no shell and one tool, manifold, on the voice run's socket, so it does exactly what the grant allows and every command is in the feed. The audio goes between your browser and the model directly (WebRTC); the core passes the handshake, the transcripts and the tool calls. Codex runs here, so this voice reaches only this machine's manifolds. Voices: juniper, maple, spruce, ember, vale, breeze, arbor, sol, cove.

Hybrid minds

manifold mind new "SPX mind" --goal "find SPX contracts worth trading when the index moves" --router gemma4:e4b --specialists sonnet,opus --watch "option_chain.spot > 7800"

A mind is a manifold where several models share one Memory, one Questions and Findings and one Brief, each sized to what it costs: gates watch for free, a router on a local model is woken with what changed and decides — a small thing itself, or ask sonnet "…" with a task it writes — frontier specialists do the work and leave it on the nodes, and the Judge is System One for a probability. Nothing runs on a timer.

Finding things

Agents search, never browse. find <words> answers in one short ranked list across a manifold's cards, its types, its docs, the node library and the machine's tools, each line with the command that uses it. library alone is a map of the domains that fit the manifold's purpose (System One judges that once, when a key is held); library <words> and tools <words> search; tools <name> shows one tool. Measured on 27 plain asks: the right type lands in the top five 96% of the time, in about 400 tokens, where a listing cost up to 39,000.

Reading the web

Agents have no browser and no network of their own: a link in a card is readable through a Reader card. An agent (or you) places one — place web/reader <section> url=https://… — you approve the type once (it reaches the web), and cat gives the page as plain markdown: title, text, links. do <node> fetch url=… reads another page into the same card. The web tool behind it reads public pages only, never anything on your machine or network, and every page read is in the feed. library <words> lets a run find a type in the node library instead of guessing its name.

Other machines

A manifold lives on one machine. Any other machine opens it by invitation, one manifold at a time, at a level:

On the canvas: Activity → Share, press a level, Copy; the Join card on the other machine's home page takes the code. A level-0 code works once and for ten minutes; a person's once and for 24 hours. The same from a terminal:

manifold -m desk share --level 0        # your other machine; there: manifold join <code>
manifold -m desk share --level 2        # a friend as Operator (0 Owner · 1 Builder · 2 Operator · 3 Viewer), within --hours, for --people
manifold join <code>                    # on the other machine; then manifold -m desk ls, or pick it on the canvas
manifold -m desk members                # who is in, at which level; members rm <name> ends one; members allow <name> <section> opens a private section

A machine of yours works in that manifold as you do, private sections and every command, and sees nothing else of the machine it lives on; both machines run their own Manifold. A friend never sees a private section until you allow it by name, never your keys, and never another manifold. Your agents and your voice run on the machine you are at and reach the manifold over its link. private is granted to a session bound to one manifold, never to one that reaches several.

A canvas on any machine shows what that machine's agents read: the docs, the cards, the library and who is here all come from the machine the manifold lives on, computed once, the way an agent's commands are answered there. You write a manifold's docs from any machine of yours; a friend reads them.

Where things are

The package contains code, the node library with synthetic sample data, and the door binary; nothing from the machine it was built on. Everything you make lives in your home folder. A key goes to the macOS keychain, never to a file. Nothing leaves your machine except what you grant to a machine you invited, through the door.

Adding a library or a service

A tool is how cards reach a library or a service. An agent drafts one — write tool/<name>/tool.json, tool.py, a doc — check tool/<name> reads it over, and you accept it on the canvas (Nodes → drafts) or with accept tool/<name>: that installs its packages in an environment of its own and makes it a tool of the machine. Then cards use it. The first one shipped this way is Manim: manifold tool add library/tools/manim, then place math_visual/manim_scene. A card that needs an API key says so and shows a field with the key's name: paste the value, save, it gathers.

Building a node

A node is three files. node.json says what it answers, its settings and the shape of its value. node.py has gather(m, s), which returns the value, and act_<name>(m, s, …) for each action. view.html draws the card. manifold bench <folder> --save --shot runs one without a core. library/GUIDE.md is how library nodes are built.

Development

git clone … && cd manifold && uv venv --python 3.13 && uv pip install -e ".[dev]"
(cd door && cargo build --release)          # the door; without it the Python door runs
.venv/bin/python -m pytest -q

DECISIONS.md holds every design decision with its reason and what would prove it wrong. AUDIT.md records what a hostile machine was tried against and what remains unproven. Apache-2.0.