W&B HiveMind

W&B HiveMind is a shared dashboard for AI coding sessions. A lightweight daemon runs on each developer's machine, watches for coding agent activity, and sends session transcripts to hivemind.wandb.tools.

Coding agents like Claude Code, Codex, Cursor, Gemini CLI, OpenCode, GitHub Copilot CLI, and Pi all write transcripts to the local filesystem as they work. The HiveMind daemon wakes up every 30 seconds, checks for new activity, and uploads anything it finds. There's nothing to configure per agent. If the agent writes transcripts, HiveMind picks them up.

This repository hosts release binaries for the HiveMind daemon and is the public issue tracker. Found a bug or have a feature request? Open an issue. The dashboard and docs live at hivemind.wandb.tools.

Getting started

Install the client and start the daemon. hivemind start registers a background service (launchd on macOS, systemd on Linux) so the daemon keeps running and starts on login, prompting you to authenticate through GitHub or your organization's SSO if you haven't already.

Install (macOS or Linux)

curl -fsSL https://hivemind.wandb.tools/install | sh
hivemind start

The one-line installer is the recommended path. On an Apple Silicon Mac with Homebrew it installs the wandb-hivemind cask, so you get a clean brew uninstall later; everywhere else it downloads the signed binary to ~/.local/bin — no Homebrew, Python, or sudo required. Either way the daemon detects and applies upgrades automatically.

To skip the cask and always install the raw signed binary — even on a Homebrew Mac — pass --binary:

curl -fsSL https://hivemind.wandb.tools/install | sh -s -- --binary

On macOS the binary requires Apple Silicon; on an Intel Mac, use uv (below). For fleet rollouts through an MDM, see Deploying with MDM.

macOS (Homebrew cask)

Prefer to drive the install through Homebrew yourself? Install the cask directly:

brew install wandb/taps/wandb-hivemind
hivemind start

Homebrew 6 requires third-party taps to be trusted before their code runs. Installing by fully-qualified name records that trust automatically (you'll see Trusted cask wandb/taps/wandb-hivemind on first install); review or revoke it later with brew trust / brew untrust.

The cask installs the same self-updating binary, so there's nothing to upgrade by hand — brew upgrade is a no-op for it unless you pass --greedy. The cask is Apple Silicon-only; on an Intel Mac, use uv below.

Any platform (uv)

uv tool install wandb-hivemind
hivemind start

uv installs the cross-platform Python package — the path for Intel Macs and Windows, where the one-line installer and cask don't apply. Upgrade with:

uv tool upgrade wandb-hivemind

Uninstalling

How you remove HiveMind depends on how it was installed. Since the one-line installer uses the cask on an Apple Silicon Mac with Homebrew, check whether brew owns it:

brew list wandb/taps/wandb-hivemind

If that succeeds, uninstall through brew with --zap — this stops and unregisters the launchd service, removes the binary, and clears the leftover LaunchAgent plist and logs (it deliberately leaves ~/.hivemind alone — see below):

brew uninstall --zap wandb/taps/wandb-hivemind

Installed with uv? Use uv tool uninstall wandb-hivemind. Otherwise it's a binary install — stop and unregister the daemon, then remove the binary:

hivemind stop --disable
rm ~/.local/bin/hivemind

Either way, your data in ~/.hivemind/ is left in place. It holds the daemon's sync state — which sessions have already been uploaded — so removing it and reinstalling later triggers a re-sync. Delete it only if you don't intend to use HiveMind again:

rm -rf ~/.hivemind

Docker (sidecar)

For containerized agents, run the daemon as a sidecar that watches the agent's transcript directory. Images are published for linux/amd64 and linux/arm64:

docker run -d \
  -v claude-sessions:/watch/.claude:ro \
  -e HIVEMIND_TOKEN=<your-token> \
  -e HIVEMIND_WATCH_PATHS=/watch/.claude \
  ghcr.io/wandb/hivemind:latest

Mount the directory your agent writes transcripts to (read-only is fine) and point HIVEMIND_WATCH_PATHS at it. The HiveMind server image is also available as ghcr.io/wandb/hivemind-server.

How it works

Once the daemon is running, open any supported coding agent and start working. Within 30 seconds your session appears on the dashboard. You can watch sessions in real-time, review past conversations, and dig into individual tool calls.

Supported agents

The HiveMind agent

HiveMind also installs an @hivemind agent definition when the daemon starts (Claude Code, Codex, and Cursor). Type @hivemind in a supported agent to ask questions about past coding sessions: what you worked on last week, how a bug was fixed, where a particular change was made. It searches across your team's session history and pulls the answer into your current conversation.

How HiveMind compares

HiveMind gives you one view across every coding agent your team uses, instead of a separate dashboard per vendor. Native dashboards from individual agent vendors only show their own usage. HiveMind brings Claude Code, Codex, Cursor, Gemini, Copilot, and more together with session-level detail, spend, and outcomes in one place.

HiveMind and Weave

If you already use W&B Weave, it works together with HiveMind. They cover different stages and answer different questions.

• Weave observes what your AI application does in production, tracking LLM and agent traces, evaluations, and quality.
• HiveMind observes how your team builds software with AI coding agents, tracking details like sessions, spend, and productivity.

Next steps

• hivemind.wandb.tools: Sign in and see your team's sessions on the live dashboard
• Documentation: Full docs, including configuration and MDM deployment