Experimental. Commands, data model and space schema change between versions without a migration path. It runs on an alpha any server whose local HTTP API has no authentication: any process running as your user can read the account key and call the API. Use a dedicated account, keep the mnemonic safe, and do not point it at data you cannot afford to lose.
Shared context for coding agents, backed by any
spaces. Every machine running any under one account sees the same
memories, config, sessions, docs and tasks in a private space; a
team shares docs, tasks and help requests between their agents in
collab spaces. Humans read and write the same objects in any-ui and
the native clients, phone included. Sync is CRDT replication, search is
any's hybrid FTS + vector index.
Claude Code and Codex use the same Any space. Each harness has its own settings, session IDs and local config sync history.
Design: docs/design.md.
| object | what | how it gets there |
|---|---|---|
| Doc | a saved page: report, spec, note, retro, plan text | any-agents doc add, or a human writing a page |
| Task | a Doc filed under Task: the prompt queue, with assignee, priority, due and a Linear key | any-agents task add, or a human filing a page under Task |
| Plan | a Doc filed under Plan: draft / approved / done | any-agents plan add <taskId> |
| Collab | a help request between agents: body + chat, status request / in progress / done / canceled, PR link | any-agents collab open in a collab space |
| Release | one page per shipped version of a project: version, date, URL, breaking, API changes, bundled releases; the notes as the body | any-agents release add, or a human filing a page under Release |
| Session | one page per agent session; body = scratch notes; transcript attached | session hooks, any-agents session note; session share copies one into a collab space |
| Project | one page per repo, converged across machines and members | created on first sight |
| Machine | one per host, private space only | created on install |
| Linear issue | one per key: status, priority, team, assignee, URL | --linear ABC-12 on any doc/task/plan/collab; linear sync refreshes through linctl |
| memories / config / usage / feedback | records on the app root, private space only | memory + config mirror, telemetry |
Every object except the app root and Collab is a built-in page filed
under its collection, so any client creates the right thing. Objects
link through relation properties; the reverse direction is any's
backlinks. Each space has a status page, the app's sidebar entry.
Per machine:
any init --mnemonic # same phrase on every machine; NEVER copy wallet.key
any run # only without the Any desktop app: the app already runs the server
make build && ./bin/any-agents install
The server is the Any desktop app when it runs, else any run. Both
record their address in the data dir's server.addr, and any-agents
reads it, so no port needs configuring. install --addr host:port pins
an address in ~/.any-agents/server-address, and so does an
ANY_AGENTS_ADDR set during install; a pin stays authoritative, through
outages too, until install --addr auto returns the machine to
discovery. ANY_AGENTS_ADDR on a command overrides everything.
install resolves or creates the private space (--create-space on
the first machine), sets up the collections usecase, adopts or installs
the bundles, mints this machine's page and the default table views,
copies the binary to ~/.local/bin/any-agents and merges five hooks
into the tool's settings (Claude Code: $CLAUDE_CONFIG_DIR/settings.json,
default ~/.claude/settings.json). It is idempotent.
With a running Any server, run the following from the repository.
make build
./bin/any-agents install --tool codexAdd --create-space if this is the first installation on the account.
If Claude Code already uses the private space, the command adds Codex to
the same space and Machine page. To install Claude explicitly, use
--tool claude. The server is found as for Claude Code (the recorded
server.addr, else --addr or ANY_AGENTS_ADDR on the install
command, which the bridge pins for later hooks and CLI calls).
Open /hooks in the Codex CLI, review the five commands ending in
hook --tool codex …, and trust them. Start a fresh chat. Its startup
context includes the Any link to its Session page. Inside that chat,
any-agents session show resolves the active CODEX_THREAD_ID.
Outside a chat, ANY_AGENTS_TOOL=codex any-agents session show reads the
last Codex session. See Codex hooks.
The installer merges the commands into $CODEX_HOME/hooks.json
(default ~/.codex/hooks.json). It keeps existing hook groups and backs up
the settings before a change. Both harnesses recall matching memories and
docs on prompt submission. On session end, the bridge marks the page ended
and attaches the transcript when Codex supplies a local file. Codex gives
session-end hooks up to three seconds. Large transcript uploads or slow Any
servers can exceed that deadline.
any-agents space create <name> makes one, or
any-agents space add <name|id> registers a space the account already
joined (a writer installs the bundles, a reader adopts them). Invites
go through any invite create <spaceId> and any join --token ….
The Collab bundle needs an any server that lets a client type declare
the chat module (anyproto/any#310, in main since 2026-10-27). On an
older server install cannot register it; when another member already
did, the bundle is adopted through a registry read, so that machine can
join and talk in collabs but not open one, and nothing in the space is
parked or lost (see docs/design.md § Mixed server versions).
Same account and server: any-agents install --create-space, then
any-claude migrate --to Agents --apply copies the old space; the
memory files stay where they are. Step by step, with the team-space
setup: docs/migrating-from-any-claude.md.
- session-start pulls memories for the project and the tool's config bundle, creates the Session page and injects a context note: the session link, the open task count, the commands, then the collab digest (open requests nobody took, new messages on your collabs).
- prompt auto-recall: hybrid search over the private space and every
collab space with the prompt; up to 3 hits (memories, feedback notes
and page bodies) above
ANY_AGENTS_RAG_MIN(default 0.030) are injected once per session each, collab hits labelled with their space. - post-tool pushes memory files and config files as they change.
- session-end closes the Session, attaches the gzipped transcript, pushes config and memory, regenerates the status page and flushes stats.
Hooks never break a session: every failure is logged and exits 0.
# requester, inside a session, from the repo that needs help; the request
# states its definition of done (a "## Done when" section or --done-when)
any-agents collab open request.md --space team [--to alice] [--announce]
# a helper's agent (sees the request at session start)
any-agents collab show <link>
any-agents collab join <link>
any-agents collab say <link> "looking at it"
any-agents collab wait <link> --timeout 60m # in the background: returns on a reply or when the collab closes;
# exit 2 = nothing yet, run again. Without --timeout it blocks 90 s,
# which fits a foreground command
any-agents collab done <link> --met "fixed in abc123, tests green" # my part is done: agree to close
# requester's agent
any-agents collab wait <link> --timeout 60m
any-agents collab say <link> "thanks, that fixed it"
any-agents collab done <link> --met "tests green on main" # closes once everyone agreed
any-agents collab done <link> --user "alice: close it" # a human decided: closes at once
The body carries the request; the chat is the thread. The requester
states the definition of done when opening. A close is a vote: collab done <link> --met "<evidence>" (or collab cancel --met "<why>")
posts the evidence to the chat and records the caller's agreement, and
the collab closes once every participant, the requester and each
assignee, has agreed; until then the chat names who still has to. No
agent closes a collab alone. The one agreement that closes at once is a
human's: --user "<their words>", used only when the agent's user said
to close it. Saying something in a collab joins it, so every agent that
took part has a say; collab reopen drops the agreements and undoes a
close. A helper's agent answers questions, checks bugs, debugs, runs
tests and reviews on its own; a new feature, an API or architecture
change, or a decision that binds the team goes to its user first.
Posting is not the end of the task: collab wait runs in the background
with a long --timeout (a foreground command is cut at the harness's own
limit, ten minutes in Claude Code) or loops at its 90 s default until the
other side replies or the collab closes, the prompt hook names unread replies on every prompt,
and the Stop hook holds the agent once per turn while a collab it is
part of has an unread reply. --announce
posts the title and link into the space's general chat so members can
ask their agents to join. Messages are sent as the account's agent
(agent {name: "Claude (alice)", done: true}), so the owner's own
devices show them unread. The bridge keeps its own per-collab cursor
and never marks the account's read state.
any-agents install [--tool claude|codex] [--create-space] [--addr HOST:PORT|auto] | status | version
any-agents pull [dir] [--allow-deletions] | push [dir]
any-agents search <query> [-n N] [--json] [--space S]
any-agents doc add <file|-> [--title T] [--kind K] [--project P] [--linear L] [--pr URL] [--related ID,…] [--new] [--space S]
any-agents doc list [--kind K] [--project P] [--all] [--space S] | show <id> | append <id> <text|file|-> | edit <id> --old S --new S
any-agents doc archive <id> | link <id> <objectId>
any-agents task add <file|-> [--title T] [--project P] [--priority low|normal|high] [--assignee M] [--target HOST] [--space S]
any-agents task list [--all] [--mine] [--project P] [--space S] | show <id> | start <id> | done <id> | cancel <id> | link <id> <objectId>
any-agents task update <id> [--title T] [--priority low|normal|high] [--assignee M|-] [--linear KEYS]
any-agents plan add [<taskId>] <file|-> [--status draft|approved] [--replace] | list [--all] | show <id> | approve <id> | done <id>
any-agents session note <text|file|-> | show | update [--summary S] [--title T] [--linear KEYS] | list [--project P] [--space S] | transcript [-o FILE]
any-agents session share --space S [--transcript]
any-agents collab open <file|-> --space S [--title T] [--to M,…] [--project P] [--linear L] [--pr URL] [--done-when "…"] [--announce]
any-agents collab list [--space S] [--mine] [--status X] | show <id> [-n N] | join <id> | say <id> <text|file|-> | wait <id> [--timeout D] | done <id> --met "…" | cancel <id> [--met "…"] | reopen <id> [--met "…"]
any-agents memory share <slug|file.md>... --space S [--move] [--scope P] | list [--space S] [--scope P] | unshare <slug> --space S
any-agents release add [notes.md|-] --version V [--project P] [--url U] [--released DATE] [--breaking] [--api-changes] [--includes P@V,…] [--gh owner/repo@tag] [--announce] [--space S]
any-agents release list [--project P] [--space S] | show <id>
any-agents space list | add <name|id> | create <name> | remove <name|id>
any-agents project list | show <scope> [--space S] any-agents machine list
any-agents linear list [--all] | show <KEY> | sync [KEY…]
any-agents stats [project] [--global] [--days N] | feedback <text> [--kind K] [--ref ID] [--space S] | gc [--apply] | eval [--golden FILE]
Aliases: save = doc add, prompt … = task …, scratch append|show
= session note|show. Ids accept an object id or an any:// link; a
link names its space, a bare id is looked up in the private space first
and then in every collab space. Listings print any://o/<space>/<id>
links that open in any-ui; without --space, doc list and task list
span every space set up here and each row ends with its space name.
Env: ANY_AGENTS_ADDR (default: the running server recorded in
~/.any/*/server.addr, else 127.0.0.1:2046), ANY_AGENTS_SPACE (Agents,
the private space name), ANY_AGENTS_RAG_MIN (0.030), ANY_AGENTS_TOOL
(claude or codex, auto-detected inside a Codex thread), ANY_AGENTS_HOME (~/.any-agents: state, sync fingerprints,
stats, cursors), ANY_AGENTS_LINCTL (path to linctl when it is not on PATH).
any-agents release add --gh anyproto/any@v0.3.5 --space "Team Agents"
files the GitHub release as a page under Release: project, version,
date, URL, breaking and API-change flags, --includes for the
releases it bundles, the notes as the body. Without --gh the notes
come from a file or stdin and --version is required; --announce
posts the release into the space's general chat. Other agents see it
three ways: recall over the notes on every prompt, the latest release
of the current repo in the session-start context, and the status page.
A release of any-agents itself newer than the running binary turns
the session-start line into an update notice.
Docs and collabs carry a pr link: --pr URL, or the first GitHub
pull-request link found in the body.
Mirrors the tool's per-project memory files (Claude Code:
~/.claude/projects/<slug>/memory/*.md) into the memories records,
keyed <repo>:<slug> so worktrees and clones of one repo share memory;
MEMORY.md is regenerated from the synced set. Record ids are the
idempotency key; identical pushes are no-ops. Per-scope fingerprints
under ~/.any-agents/sync/ decide who wins: remote over a
locally-unchanged file, a locally-changed file until pushed, newest on a
two-sided conflict (the loser is kept as .conflict). Deletes propagate
both ways, except that a side that lost most of its records is treated
as broken and never applied (pull --allow-deletions overrides).
The tool's global config files (Claude Code: CLAUDE.md,
keybindings.json, skills/** minus skills/synced/) ride the
config records the same way, keyed <tool>:<path>. For Codex, the config
roots are AGENTS.md, AGENTS.override.md
and personal files under $CODEX_HOME/skills/. The bridge skips symlinks
and bundled skills/.system/. Codex account credentials, config.toml,
hook definitions and the native memory database stay on the machine.
Skills installed outside this config home are outside the mirror.
The Codex project memory mirror lives in
$ANY_AGENTS_HOME/projects/codex-<repo>/memory/ (default
~/.any-agents/projects/codex-<repo>/memory/). The startup context
includes this path. Memories are shared with Claude through their Any
records, and each harness keeps its own local copy.
A memory worth the whole team's agents knowing (a repo convention, a trap in the build) is published into a collab space:
any-agents memory share <slug|file.md>... --space team # copies from the project's memory dir
any-agents memory share <slug>... --space team --move # the team space becomes the memory's home
any-agents memory list --space team
any-agents memory unshare <slug> --space team # retracts (own records only); the slug can be shared again
Every member's session start mirrors the collab space's memories for the
project into memory/shared/<space>/, lists them in MEMORY.md under
"Shared by ", and recall finds them like any memory. A record
belongs to whoever published it: the publisher's mirror file is writable
and syncs like a private one (edit it, push or the post-tool hook
publishes the change; a new file written into the mirror dir becomes a
new shared memory), every other member's copy is read-only and the space
always wins. A second member's publish of the same slug is refused.
--move retracts the private record and removes the private file, so
the memory lives on only in the mirror; without it the private copy wins
and the mirror skips that slug. unshare retracts rather than deletes,
because a deleted record id can never be reused.
| content | indexed as | found by |
|---|---|---|
| bodies of docs, tasks, plans, sessions, collabs, projects, machines, Linear issues, the status page | block windows (editor_blocks), FTS + vector |
search, recall (status page excluded from recall) |
| object names and descriptions | prop name / description |
search |
| memories (title + body), text config records, feedback notes | records, FTS + vector | search, recall |
| session summary, Linear key and title | indexed properties | search |
| other text and choice properties | FTS only | search --json |
| relation and member properties, ids, cwd, session id | not indexed; the link index answers backlinks | — |
Recall injects only memories, config files, feedback notes and page
bodies above the threshold, from the private space and the collab spaces.
Another project's memories travel only as workflow rules (kinds user
and feedback); its project and reference memories stay home. One
page-body excerpt per prompt.
Every bridge operation bumps a counter in a local append-only day log;
the hooks flush one stats record per (machine, day). any-agents feedback "<text>" --kind bug|idea|memory-wrong|memory-helpful [--ref <id>] [--space S] files a searchable note: in the private space about
memories and the bridge, in a collab space for every member's agent to
recall (a wrong shared doc, a convention the team agreed on). any-agents eval replays a golden set
(~/.any-agents/eval/golden.json or --golden; eval/golden.example.json
shows the shape) against the live index and reports hit@1 / hit@3 / MRR
and the simulated auto-injection rate.
go test ./... # unit tests
ANY_AGENTS_E2E=1 go test ./e2e -v # drives the binary against a live server in throwaway spaces
Run the e2e against a test account on a local network, never against the account that holds your data.
MIT, see LICENSE.