@edspencer/managers
v0.5.0
Published
Scheduled per-project manager agents that track long-running objectives. A fork of Paddock, built on herdctl.
Maintainers
Readme
Managers
Scheduled, per-project manager agents that keep track of your long-running objectives. Each project gets a manager: a Claude Code agent that wakes on a schedule, reads where things stand, works on the project's objectives, and writes down what it did and what it needs from you. Its state is plain Markdown and YAML in a git-tracked data directory. A fork of Paddock, built on herdctl.
npx @edspencer/managersThen open http://127.0.0.1:7234.
First run downloads ~250 MB. Managers drives Claude Code, and the Claude Agent SDK ships a per-platform binary of that size. It cannot be skipped: installing with
--omit=optionalproduces a Managers whose chats all fail. Later runs reuse the npm cache. For repeated use,npm i -g @edspencer/managersand thenmanagersis friendlier than barenpx.
Real turns spend real Claude credit. Every behaviour (scheduled wakes, reports, memory consolidation) is off until you switch it on in the UI.
Credentials
Managers runs Claude Code on your behalf, so it needs Claude credentials. If you
already use Claude Code on this machine, there is nothing to do: it uses the
login you already have (the macOS Keychain entry on a Mac, or
~/.claude/.credentials.json elsewhere), read, never copied. Otherwise:
claude setup-token # Claude Max/Pro
export ANTHROPIC_API_KEY=sk-ant-… # or API billingManagers keeps its own Claude home under the data dir, so its transcripts never mix with yours.
Security
The managers command binds 127.0.0.1 and runs with authentication off,
so nothing off this machine can reach it. Processes on this machine can, and
that includes the agents Managers runs: with no auth, an agent's Bash can call
the API as you. The UI shows a permanent banner saying so.
- To close that, use
MANAGERS_AUTH_MODE=jwtbehind your identity provider (see AUTH.md). - To make the command refuse to start without auth, set
MANAGERS_DANGEROUSLY_ALLOW_NO_AUTH=0. - It refuses to start if you bind a routable interface (
--host 0.0.0.0) with auth off. ConfigureMANAGERS_AUTH_MODEfirst.
Options
-p, --port <port> HTTP/WS port (default 7234)
--host <host> Bind address (default 127.0.0.1)
-d, --data-dir <path> Projects + state (default ~/.managers)
-o, --open Open the app in your browser once it is listening
-q, --quiet Print only the first line, `Managers <version>: <url>`
--verbose Show the server's own logs (quiet by default)
-v, --version Print the version
-h, --help Show helpmanagers service install keeps it running in the background from login
(launchd on macOS, a systemd user unit on Linux; it refuses to install from the
npx cache or a git checkout, so npm i -g first, and refuses a port something
else holds). Running it again keeps the port, host and data dir the installed unit
records unless you give new ones; --dry-run prints the unit and writes nothing,
and --json on install, uninstall, start, stop and restart prints one
JSON document. managers config show --resolved prints every
effective setting and where it came from.
Your projects, objectives, tasks, memory, chats and settings persist in
~/.managers, or wherever --data-dir points. It is one directory, and the
projects inside it are a git repository the managers commit to. Move it to move
your instance; delete it to start over.
Requirements
Node.js 22 or newer.
Other ways to run it
A multi-arch (amd64 + arm64) Docker image is published alongside this package, and is the better fit for a server, a VM or an LXC:
docker run -d --name managers -p 127.0.0.1:7234:7234 -v managers-data:/data \
-e CLAUDE_CODE_OAUTH_TOKEN=… \
-e MANAGERS_DANGEROUSLY_ALLOW_NO_AUTH=1 \
-e MANAGERS_DANGEROUSLY_ALLOW_OPEN=1 \
ghcr.io/edspencer/managers:latestThe image binds 0.0.0.0 inside the container, so with auth off it needs both
opt-ins, and publishing on 127.0.0.1: is what keeps that private. For anything
reachable from a network, use MANAGERS_AUTH_MODE=jwt instead (see AUTH.md).
