Context Tree gives coding agents lasting project memory: decisions, constraints, and the reasons behind them. Use it to avoid repeating context in new sessions, keep different agents aligned, or share knowledge across related repositories. Context lives in a separate Git repository. Keep it local or share it through private GitHub storage.

Install

Requires Node.js 22.13+ and Git. For GitHub sharing, also install and sign in to the GitHub CLI (gh auth login).

npm install --global @first-tree-ai/context-tree

This installs the CLI and skills for installed Codex, Claude Code, and Pi agents. Restart your agent to discover the skills.

Get started

Open your project in your agent and ask:

Set up a local Context Tree for this project, then read the relevant context.

You can also ask to connect an existing tree or create a private GitHub tree. To create a local tree yourself, run this from your project directory:

context-tree create

The tree is saved under ~/.context-tree/trees and connected to your project.

Read and save context

Ask your agent to use the tree as you work:

Read the Context Tree before planning this change.

Save our decision to use a single writer, including why we rejected multiple writers.

The read skill retrieves relevant context; the write skill updates it and commits changes, pushing them when the tree is shared on GitHub. Save decisions and constraints that future work should respect, with their rationale.

To invoke a skill explicitly, use $context-tree-read in Codex, /context-tree-read in Claude Code, or /skill:context-tree-read in Pi. Replace read with write, setup, create, connect, publish, cleanup, or schedule-cleanup for the other workflows.

Publish your local tree as a new private GitHub repository:

context-tree publish OWNER/REPO

From another project or machine, connect to it:

context-tree connect OWNER/REPO

You can also reuse a local tree by name or connect a checkout on disk:

context-tree list
context-tree connect my-project-context-tree
context-tree connect --tree-path /absolute/path/to/tree

Connecting switches the current project's tree. Run these commands from the project directory, or add --project-path /path/to/project.

Cleanup and scheduling

Ask your agent to remove outdated clutter and consolidate duplicate context:

Run the context-tree-cleanup skill for this project and publish the changes.

For recurring cleanup, use the schedule-cleanup skill or run:

context-tree cleanup schedule --agent codex --every 2h
context-tree cleanup status
context-tree cleanup logs
context-tree cleanup remove

Choose codex, claude, or pi; the agent CLI must be installed and authenticated. Schedules run locally on macOS or Linux while the machine is awake, and skip trees unused for 24 hours. Use one designated cleaner per shared tree. context-tree cleanup run runs the configured cleanup now; cleanup logs --list lists previous runs.

Useful commands

context-tree resolve                              # show this project's tree
context-tree read --tree-path /path/to/tree        # browse its root index
context-tree verify --tree-path /path/to/tree      # check its structure
context-tree install                              # add skills for a new agent
context-tree install --project .                  # install skills for this project
context-tree uninstall                            # remove Context Tree skills
context-tree --help

For scripts and custom integrations, see the CLI specification, including synchronization, prepared writes, and JSON output.

Working with OpenTag

Create or publish a tree as above, then choose it on each OpenTag Computer:

opentag context-tree connect OWNER/REPO
# Or use a local tree:
opentag context-tree connect my-project-context-tree

OpenTag connects each agent workspace to that tree when a session starts. Use the same read, write, and cleanup prompts in those sessions. For a GitHub tree, GitHub authentication must work on each Computer.