Tendril

Tendril is an agentic orchestration tool that parallelizes software development across multiple AI agents. It uses git worktrees to create isolated working directories — each on its own branch with Claude Code ready to go — so you can run multiple AI-assisted dev tasks simultaneously without leaving your host machine.

Instead of running one agent on one branch, Tendril lets you orchestrate many agents at once: open a terminal, run tndrl start, and fire off a task. Open another terminal, do it again. Each agent works independently in its own worktree, on its own branch, with full repo access and all your existing dev tools.

Prerequisites

• macOS (v1 target platform)
• Rust 1.85+ (uses edition 2024)
• Git with an origin remote configured
• Claude Code CLI installed (npm install -g @anthropic-ai/claude-code)

Install

git clone <this-repo>
cd tendril
cargo build --release

The binary is at target/release/tndrl. Add it to your PATH or use cargo install --path ..

Usage

Start a new session

tndrl start

This will:

1. Detect your repo's git remote URL from the current directory
2. Prompt you for a branch name
3. Create a git worktree at ~/.tendril/worktrees/<session-id> (creates the branch if it doesn't exist)
4. Launch Claude Code directly in the worktree directory

After launching, you'll land in a Claude Code session inside the worktree. Open another terminal and run tndrl start again to spin up a parallel session on a different branch.

You can also pass the branch name directly:

tndrl start --branch feature/my-task

List active sessions

tndrl list

Displays a table of all sessions with their status, branch, repository, and start time.

+----------+---------+----------------+----------------------------------+------------------------+
| ID       | Status  | Branch         | Repository                       | Started At             |
+----------+---------+----------------+----------------------------------+------------------------+
| a1b2c3d4 | Running | feature/auth   | git@github.com:user/repo.git     | 2026-02-13 15:30:00 UTC|
| e5f6g7h8 | Stopped | feature/api    | git@github.com:user/repo.git     | 2026-02-13 14:00:00 UTC|
+----------+---------+----------------+----------------------------------+------------------------+

Watch mode

tndrl list --watch

Continuously monitors session state and updates the table in real time.

Prune stopped sessions

tndrl prune

Removes stopped and failed sessions, cleaning up their worktree directories.

Nuke everything

tndrl nuke

When worktrees and sessions.json get into a bad state, nuke provides a clean slate. It scans both sessions.json and the ~/.tendril/worktrees/ directory (catching orphaned directories too), lists every worktree and its branch, then asks for confirmation before removing everything and clearing the session file.

Sprint mode (Azure DevOps)

Automatically create sessions for every work item in your current ADO sprint:

tndrl sprint --org myorg --project MyProject

This will:

1. Fetch the current sprint/iteration from Azure DevOps
2. Pull all active work items (filtering out Closed/Removed/Done)
3. Create a worktree and branch for each (e.g. feature/12345/add-user-login)
4. Open a new Terminal.app window per work item with Claude Code pre-loaded with the work item context

Configuration

Sprint mode requires a Personal Access Token with vso.work scope. Set it as an environment variable:

export ADO_PAT="your-personal-access-token"

Organization, project, and team can be passed as flags or environment variables:

CLI flags take precedence over environment variables.

Options

# Preview sprint items without starting sessions
tndrl sprint --org myorg --project MyProject --dry-run

# Run Claude in the background instead of opening Terminal windows
tndrl sprint --org myorg --project MyProject --headless

# Use env vars so you don't need flags every time
export ADO_ORG=myorg
export ADO_PROJECT=MyProject
tndrl sprint

In headless mode, each Claude process runs in the background with output logged to claude.log in the worktree directory. Session status is tracked via PID liveness — tndrl list will detect when background processes finish.

How it works

Git worktrees let you check out multiple branches of the same repository simultaneously in separate directories. Tendril leverages this to create isolated workspaces without any virtualization overhead.

Each worktree session:

• Creates a new directory at ~/.tendril/worktrees/<session-id>
• Checks out the specified branch (or creates it if it doesn't exist)
• Inherits your host's SSH keys, git config, dev tools, and environment
• Runs Claude Code natively — no overhead
• Tracks session state in ~/.tendril/sessions.json

This is fast, lightweight, and leverages everything already installed on your machine.

Session tracking

Session metadata is persisted to ~/.tendril/sessions.json. The list command reads this file and reconciles with the actual state by checking worktree directory existence.

Project structure

src/
  main.rs      -- CLI parsing, start/list/prune/nuke/sprint command orchestration
  ado.rs       -- Azure DevOps API interaction (sprint fetch, work item details)
  worktree.rs  -- Git worktree creation, branch detection, repo URL, Claude launch
  session.rs   -- Session persistence to ~/.tendril/sessions.json
  types.rs     -- Session, SessionStatus, AdoWorkItemRef types

Current limitations

• No tndrl stop command — worktrees persist until manually cleaned up
• SSH remotes preferred — HTTPS remotes with credential helpers aren't handled yet
• No session cleanup — stopped sessions accumulate in the JSON file until pruned
• Branch switching in worktrees is not tracked — Tendril assumes a 1:1:1 mapping between session, worktree, and branch. If you switch branches inside a worktree (e.g. git checkout -b new-branch), session metadata becomes stale: tndrl list will show the old branch, and tndrl start --branch new-branch will fail because git only allows a branch to be checked out in one worktree at a time. Avoid switching branches inside a Tendril worktree — start a new session instead.
• Prune doesn't check for active use — tndrl prune will remove worktree directories for stopped sessions even if your shell is still in that directory or you've started new work there. Make sure you've exited the worktree before pruning.

Tech stack

• Rust (edition 2024)
• clap — CLI argument parsing
• tokio — async runtime
• dialoguer — interactive terminal prompts
• comfy-table — formatted table output
• serde/serde_json — session serialization
• chrono — timestamps
• uuid — session IDs
• dirs — home directory resolution
• colored — terminal colors
• reqwest — HTTP client (ADO API)
• html2text — HTML-to-plain-text conversion for work item fields

Roadmap

• tndrl stop command with worktree cleanup
• Session descriptions and filtering
• Reconcile worktree branch state on list and prune (detect branch drift via git branch --show-current)
• Guard prune against active worktree usage (check for running processes before removing)
• Warn when running tndrl start from inside an existing Tendril worktree