Install

any is a single binary. It runs the server (any run) and is the CLI client for everything else. The checks on this page and the quickstarts after it use curl and jq.

From a release tarball

Releases ship one tarball per platform:

Each tarball contains:

any[.exe]        # the server + CLI (built with the local embedder)
llamacpp/        # prebuilt llama.cpp shared libs for the local embedder
manifest.json    # { version, os, arch, llamacpp_version, sha256: {path: hash} }

Extract it somewhere on your PATH, keeping llamacpp/ next to the binary — the local embedder looks for it at <dir-of-any>/llamacpp by default. The embedding model itself (~600 MB) is not bundled; it downloads on first boot into <data-dir>/models/ without blocking the server.

mkdir -p "$HOME/.local/opt/any" "$HOME/.local/bin"
tar -xzf any-*-linux-x86_64.tar.gz -C "$HOME/.local/opt/any"
ln -s "$HOME/.local/opt/any/any" "$HOME/.local/bin/any"
export PATH="$HOME/.local/bin:$PATH"      # put this in your shell profile too
any version

On macOS substitute your tarball's name; on Windows add the extracted directory to PATH.

From source

Requires Git, Go 1.26.2 or newer, and make.

git clone https://github.com/anyproto/any
cd any
make build              # → bin/any  (also fetches bin/llamacpp)
export PATH="$PWD/bin:$PATH"
any version

Run the PATH line in every terminal you use below, or put the repo's bin on PATH in your shell profile: the any commands then run the binary you just built, with bin/llamacpp next to it where the local embedder looks. Otherwise a bare any fails or runs an older install. If the llama.cpp download fails, make llamacpp retries it.

On NixOS (or any system without libffi.so.8 on the loader path) run both the build and the binary through the repo's dev shell: nix develop -c make build, nix develop -c any run.

With go install

go install github.com/anyproto/any/cmd/any@latest

This binary has full-text search and can do vector search, but it has no local embedder: index.embedder: local fails at boot, and the default auto embeds only once index.openai.baseUrl / apiKey name an OpenAI-compatible provider — indexed text and search queries then go to that API. Until then the index is full-text only.

For the local embedder, install with its build tag:

go install -tags llamacpp github.com/anyproto/any/cmd/any@latest

That binary needs a loadable libffi.so.8 on Linux at startup, and the llama.cpp shared libraries of the release pinned in internal/indexer/llamacpp_release.go, either in llamacpp/ next to the binary or at index.local.libDir. A release tarball or make build ships both ready to use (Builds and CI).

Choose the data root and the network

Everything below runs in a dedicated data root with the local embedder, so an experiment never touches ~/.any/ and never sends text to an online embedding provider:

export ANY_DATA_DIR="$HOME/.any-demo"
export ANY_INDEX_EMBEDDER=local

Environment variables override the config file and flags override both (Configuration). local runs llama.cpp on this device; the ~600 MB model downloads into $ANY_DATA_DIR/models/ on first boot and the object API serves while it does. none keeps full-text search and skips the model. An agent's LLM provider is a separate setting on the anyrt side (Embedders).

Note. With nothing configured the server syncs against the production any-sync network under the account you are about to create. For a throwaway environment set ANY_NETWORK_NODECONF_PATH before any init, and keep one data root per network — see Networks.

Create an account

Same terminal, same exports:

any init

The recovery phrase prints to stderr, once, between two rules; stdout carries the result:

{
  "accountId": "A8tR…",
  "created": true
}

The wallet lands at $ANY_DATA_DIR/<accountId>/wallet.key — plain JSON unless ANY_WALLET_PASSKEY is set (Security model). A second any init on the same data dir changes nothing and lists the accounts it holds (--new adds another).

The mnemonic is the only way to restore the account on another device; there is no server-side recovery (Encryption). To add a second device later:

any init --mnemonic-stdin < phrase.txt    # same account id, fresh device key

Start the server

any run

LISTENING 127.0.0.1:7001

Leave it in the foreground; it stops on Ctrl-C, or on any stop --data-dir "$HOME/.any-demo" from a terminal without the exports. In a second terminal:

any status
curl -fsS http://127.0.0.1:7001/v1/auth | jq -e '.authorized == true'

{
  "status": "ok",
  "version": "any v0.1.2 (commit 1a2b3c4, built 2026-09-09T10:00:00Z)",
  "startedAt": "2026-09-10T08:00:00Z",
  "networkId": "N83gJpVd9MuNRZAuJLZ7LiMntTThhPc6DtzWWVjb1M3PouVU",
  "account": "A8tR…",
  "bootstrapping": false,
  "crdtVersion": { "supported": 2, "stored": 2, "newer": false }
}

The server serves as soon as it prints LISTENING; bootstrapping flips to false once the background space-loading pass finishes. The auth check prints true; false means init and run saw different data roots — a fresh root boots unauthorized until an account is created or selected (Accounts).

Where things live

Next: curl or CLI.