Skip to content
jmanuelcorralPublic

About

BOND: harness multiagente para pi, Claude Code, Copilot, Codex, opencode, Crush y Gemini, coordinado con Herdr

Resources

Stars

2 stars

Watchers

0 watching

Forks

Latest commit

 

History

386 Commits

Folders and files

Repository files navigation

bond

Your coding agents, working together. · 007 · licence to code

English · Español

CI Latest release Downloads Homebrew Linux, macOS and Windows MIT license Documentation

bond is a CLI for working with several coding agents at once: pi, Claude Code, Codex, Copilot CLI, opencode, Crush y Antigravity, coordinated with Herdr. None of them is required: one is enough.

Hub mode in action: a pi orchestrator splits a hello world between a React frontend and an Express backend, one agent per repo
New in 1.10.0, hub mode: a pi orchestrator splits a story between a frontend and a backend, one agent per repo. Full video and what's new →

In three commands

bond setup      # once per machine: Herdr, model endpoints, which agents you work with
bond init       # once per repo: which agents to integrate, the default one, which tools
bond            # every day, inside the repo: Herdr with the default agent (gru)
  • bond setup checks Herdr (and tells you how to install it if it is missing), detects Ollama and LM Studio, lets you add OpenAI or any compatible endpoint (such as your own gateway) and brings those models to the agents that can use them. You pick your agents and a default one; your instructions (~/.bond/AGENTS.md) reach all of them. Everything lives in ~/.bond/, keys with 600 permissions.
  • bond init generates each chosen agent's config in the repo (instructions, MCP servers, skills, commands) from a single .bond/bond.toml, without overwriting what you wrote by hand. With pi, also the Bond Extensions (status footer, header, repo_map, /delegate, /team).
  • bond opens the workstation in Herdr. If it already exists, it focuses it; if you are already in it, it opens the default agent in that pane. bond --agent claude, bond --worktree feat/x and bond team feat/a feat/b:claude for more agents at once.

Documentation

On the web: josecorral.dev/bond/en, in English and in Spanish (the same docs/en/ and docs/, with search). In the repo:

Doc What for
docs/en/whats-new.md What is new in each release; in 1.10.0, hub mode (with video)
docs/en/getting-started.md From zero to working: install, setup, first repo, daily loop, glossary
docs/en/workstation.md bond, bond team, bond status: workspaces, agents, worktrees, teams
docs/en/reference.md Reference: every command, what it generates for each agent and why
docs/en/troubleshooting.md Quick diagnosis (bond doctor) and common errors
CHANGELOG.md Changes per release (in Spanish)

Install

# Linux and macOS
curl -fsSL https://raw.githubusercontent.com/jmanuelcorral/bond/main/install.sh | sh

# Homebrew (macOS and Linux)
brew install jmanuelcorral/tap/bond
# Windows
powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/jmanuelcorral/bond/main/install.ps1 | iex"

The installers download the binary from the latest release and check its SHA-256. Then: bond setup. To update: bond update.

From source (Node 22.18 or later): git clone https://github.com/jmanuelcorral/bond && cd bond && npm ci --ignore-scripts && bin/bond setup.

The CLI speaks Spanish for now (messages and help); the documentation is in both languages.

Multi-agent

What How
Another agent instead of the default one bond --agent claude
Work in a worktree bond --worktree feat/x --base main (uses herdr worktree)
A team, one agent per branch bond team feat/a feat/b:claude --base main: inside Herdr, minions in panes of your tab
How each branch is going, send it work, collect bond team status, bond team prompt feat/a "…", bond team clean
See what each one is doing bond status; bond sessions to resume pi sessions
Delegate to another agent from an agent bond-delegate skill: in pi /delegate reviewer:claude review the diff; in Claude or Copilot /bond-delegate …
A hub: an operations repo with other repos (submodule or symlink) bond hub init creates it and imports the repos into repos/; bond hub check, bond hub pull and bond team fix/x (one vertical branch, one agent per repo)
Split a task across worktrees from an agent bond-team skill: in pi /team <task>; it proposes the split, waits for your OK and sums up per branch

bond-delegate and bond-team are global skills: any agent can delegate to any other. Everything follows the rules of the herdr skill: Herdr is only used when you ask for it, unique names, no automatic merge, push or cleanup, and nobody answers an agent's questions for you.

Hub mode

A hub is an operations repo that groups other repos (as submodules or as symlinks to your clones) and works on them vertically: a branch with the same name in all of them and one agent per repo, coordinated from the hub.

bond hub init                    # creates the hub, imports the repos into repos/ and runs bond init in them
bond hub check                   # what each repo is missing and how to fix it (--fix)
bond hub pull                    # brings each repo's base up to date (fast-forward only)
bond team fix/login              # a vertical branch: one worktree and one agent per repo
bond team prompt fix/login "login fails with uppercase emails"
bond team collect fix/login --repo backend && bond team clean fix/login

The session lives in hub.toml, and the hub's AGENTS.md carries a generated block with the repos and the hub's rules for the orchestrator. Details in What's new and in the reference.

Workflow with skills

bond init can install skills in the repo that take a task from the issue to the release. Each step shows what it is going to write or publish and waits for your OK:

refine the issue → plan in phases → run a phase → commit → PR → release

code-context, git-conventional-commit (used by /commit), plan-create and plan-execute come selected; issue-refine-github, delivery-review-github, release-prepare-github (all three need GitHub), convention-document and plan-team are optional. plan-team runs the independent phases of a plan in parallel, one bond team branch per phase. Details in docs/en/reference.md.

Context layer

So that the agent understands a repo by reading as little as possible: pi's native search tools (grep, find, ls), the repo_map tool (a map of the repo in ~40 lines), library docs by version (context7 MCP), web search (duckduckgo MCP), ast-grep and the code-context skill. Details and reasons in docs/en/reference.md.

pi mcp list            # check the MCP servers

Development

npm ci --ignore-scripts          # dependencies (and TypeScript to check types)
bin/bond <command>               # bond from source (Node >= 22.18)
npm run typecheck && bun test    # types and tests (test/)
bun scripts/build.mts            # dist/: Linux, macOS and Windows binaries + npm package

The binary embeds the catalog, the global skills and the Bond Extensions, and extracts them the first time into ~/.bond/runtime/<version>/. CI (.github/workflows/ci.yml) checks types, runs the tests, builds it and tests the binary on Linux, macOS and Windows. Contributions follow AGENTS.md (in Spanish).

Layout

bond/
├── bin/bond                      # entry point (checks Node and loads cli/)
├── cli/                          # the CLI (TypeScript .mts, run from the repo without building)
│   ├── main.mts                  # subcommands
│   ├── setup.mts, wizard.mts     # bond setup
│   ├── repo.mts                  # bond init, bond sync
│   ├── catalog.mts               # bond catalog (create, check, add, list, update, remove, clean)
│   ├── catalog-git.mts           # catalogs over git: refs, cache and .bond/lock.json
│   ├── launch.mts, herdr.mts     # bond, bond dev, bond status
│   ├── team.mts                  # bond team (status, prompt, collect, clean)
│   ├── hub.mts, hub-init.mts     # bond hub: the session (hub.toml), init, add, sync
│   ├── hub-check.mts, hub-pull.mts   # bond hub check and bond hub pull
│   ├── sessions.mts              # bond sessions
│   ├── decisions.mts             # bond decisions: the decision memory and its runs
│   ├── trust.mts                 # agents' trust in bond team worktrees
│   ├── update.mts                # bond update
│   ├── config.mts, keys.mts      # ~/.bond: config.toml, keys
│   ├── endpoints.mts, models.mts, agents.mts, run.mts   # endpoints and models per agent
│   ├── instructions.mts          # each agent's global instructions
│   ├── catalog-agents.mts        # known agents and how to install them
│   ├── doctor.mts, suggestions.mts, migrate.mts, runtime.mts, util.mts
├── catalog/                      # what bond init installs in a repo
│   ├── catalog.toml              # name, summary of each skill and whether it comes selected
│   ├── skills/                   # code-context, plan-create, plan-execute...
│   ├── commands/                 # review, commit, handoff (common syntax)
│   └── mcps/                     # MCP servers, one per file: context7, duckduckgo
├── templates/catalog/            # what bond catalog create writes in a new catalog
├── global/skills/                # bond-delegate, bond-team (global, for any agent)
├── assets/                       # the README's animated preview
├── extensions/, prompts/, lib/   # pi's Bond Extensions (package: package.json → "pi")
├── scripts/build.mts             # binaries and npm package
├── docs/                         # the documentation, docs/en/ in English (also on GitHub Pages)
├── site/                         # the docs/ site with VitePress (.github/workflows/pages.yml)
└── .github/workflows/ci.yml      # builds and tests on Linux, macOS and Windows

herdr/config.snippet.toml is applied by bond setup --herdr-config; scripts/pi-resources lists what pi really loads in a directory (for debugging).

License

MIT. The status footer and bond sessions start from ideas in Jeyapaul/pi (MIT).

About

BOND: harness multiagente para pi, Claude Code, Copilot, Codex, opencode, Crush y Gemini, coordinado con Herdr

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages