Your agents learn once, everywhere it applies.

01Join your workspace

Open your invitation and sign in with the invited email address, or create a workspace of your own. Your administrator assigns your role, team and clearance: they decide what your agent may read.

02Create an agent key

In your workspace, open Agent keys and create a key for this machine. Each key follows your account’s current permissions and can be revoked on its own.

03Install and connect

You need Node.js 20 or later and a signed-in Codex, Claude Code or Pi. For one project:

npm install -g @bunshin-cc/bunshin
cd ~/code/payments
bunshin setup --harness claude-code     # or codex | pi

Paste the key at the hidden prompt. Bunshin reads your permissions, saves the key privately outside the project and configures the hooks. Model calls use your agent’s own login.

04A folder of projects

To connect every project under one folder (each its own repository), run setup in the folder with --hooks user. The hooks go in your user settings, so sessions anywhere under the folder use this workspace; sessions elsewhere are untouched.

cd ~/learn
bunshin setup --harness claude-code --hooks user

bunshin sources                      # which folders take part
bunshin sources exclude narr-review  # leave one out (new folders still join)
bunshin sources only shop blog       # just these

Use one workspace per company: connect ~/dev to your company’s workspace and ~/learn to your own. Each keeps its own copy of its skills and links them only into its own projects.

Agent integrations

05Check the loop

bunshin status
bunshin sync --json

Status shows the workspace, this project’s source and whether it takes part, the library and the hooks. Teach your agent a reusable fact, end the session, and the lesson appears in Sessions and Library; a fresh session gets it when it applies.

Every day

Mostly nothing: hooks sync at session start, serve the one lesson both recall voters agree on at each prompt, push the rules attached to files before they are edited, and reflect at session end.

Slash commands

Scopes

Open Scopes in the workspace. Sources appear on their own; admins create groups and drag projects (or groups) onto them, then Save. Removing a group moves its projects and lessons up to where it sat.

1. Reflection places each lesson at the narrowest scope it holds for: this project, one of its groups, or everywhere.
2. Inside a group it is shared automatically with every project in it.
3. Each night, generalization writes what a group’s projects learned separately once into the group and trims it from the originals. Borrows count as evidence. One-person workspaces apply it at once; companies review it in Review.
4. Anyone can borrow across with /borrow.

Groups decide relevance only: who may read a lesson is still its clearance.

Review and questions

The workspace’s Review page shows every proposal with its diff, where it was learned and where it applies. On a session’s page you can tell reflection what it missed; it re-reflects with your steering.

The library

Runners

Without a model key on the server, a machine answers the nightly’s queue (reflection, merges, new documents, generalization) with its own agent login. Create a runner key in Runners:

bunshin runner install --key <runner key>   # this computer, every hour
bunshin runner --once                         # answer what is queued now
bunshin runner status

Robots

bunshin setup --robot --robot-id robot-06 --arm so101 --gripper parallel \
  --mount wristcam-v4 --cameras scene,wrist --domain cooking --site lab --cell c1

A robot is its own source: group robots in Scopes and their lessons are shared the same way.

All commands

Every command and option, with troubleshooting, is in the command reference. In a terminal: bunshin --help.