AgentBridge links Claude Code, Codex and Antigravity skills

AgentBridge is a command-line tool, published on GitHub, that points the skills directories of Claude Code, Google Antigravity and OpenAI Codex at a single hub at ~/.agentbridge/skills, making a skill added once visible to all three. It runs on Node.js 18.0.0 or newer and can be executed without installation through npx @helter/agentbridge.

At a glance

• The hub is wired into each agent's own skills directory with a POSIX symbolic link on macOS and Linux or an NTFS directory junction on Windows, without a daemon, polling or administrator rights.
• MCP servers are copied rather than shared, because a server definition sits inside a config file alongside each agent's unrelated settings, so they only match up when sync-mcp runs.
• The ten-command CLI also mirrors AGENTS.md into CLAUDE.md, GEMINI.md, CODEX.md, .cursorrules and .github/copilot-instructions.md, while doctor audits broken links, plain-text secrets and conflicting server definitions across agents.

Anyone running more than one coding agent currently maintains parallel copies of the same skills, MCP entries and rule files, and every edit has to be repeated per agent. AgentBridge's split treatment reads as an honest account of the limits: directories can be shared at the filesystem level, single fields inside a shared config file cannot. The cost is that MCP state is only as current as the last sync, which likely makes doctor the load-bearing command.

Skills live in one folder that all three agents' skills directories point into

A skill is an isolated unit on disk, a folder containing SKILL.md and nothing else, so each agent's skills directory can be a link into the same hub folder. Editing a skill under any path changes the single underlying file, and link-skills only has to run once to create the links.

link-skills merges existing skills into the hub before linking, takes a pre-link backup unless --no-backup is passed, and supports --dry-run and -y. If it fails partway through, the error names the agent that failed and the agents already processed, and points to agentbridge doctor. add-skill scaffolds a new skill in the hub with YAML frontmatter.

status prints a read-only overview of detected agents, skill counts, hub link status and MCP server counts, and accepts --json for machine-readable output. pick, aliased as import, offers an interactive multiselect for importing individual skills, commands or MCP servers into the hub or a custom target.

sync-mcp rewrites matching secrets as environment variable references before writing configs

sync-mcp reads MCP declarations from every detected config, deep-merges environment variables, command arguments and server keys, then writes the result back: JSON for Antigravity and Claude, TOML for Codex, plus a copy in the hub's own ~/.agentbridge/mcp_servers.json registry used as a merge source.

Before writing, each env value is compared against the current OS environment, and an exact match is written as a ${VAR} reference instead of the literal secret. Configs still holding a literal value get rewritten on the next run. doctor flags plain-text secrets with the same suggested fix but never sets an environment variable on its own.

When two agents define the same server with different command, args or an overlapping env key, the later-processed agent's value wins on the conflicting fields and sync-mcp prints a warning naming the server, the fields and the agents. doctor repeats it as a persistent MCP Server Conflicts check.

Windows without Developer Mode gets a hardlink instead of a symlink

sync-rules treats AGENTS.md as the source of truth and mirrors it to CLAUDE.md, GEMINI.md, CODEX.md, .cursorrules and .github/copilot-instructions.md, with --mode symlink as an alternative to copies. On Windows, real file symlinks need Developer Mode or an elevated shell; without either, AgentBridge falls back to a hardlink.

A hardlink stays current when AGENTS.md is edited in place but can go stale if the file is replaced outright, as with atomic saves; the CLI reports Symlinked to source or Hardlinked to source so the difference is visible. Skills linking itself needs no elevated privileges on any of the three platforms.

watch runs a debounced auto-sync over the hub and the workspace AGENTS.md, rollback lists and restores timestamped snapshots from ~/.agentbridge/backups, and unlink removes the links and restores standalone copies. Without --fix, any warning in the doctor report exits 1, so it can gate a deploy step.

Cursor and Copilot remain rules-only

Cursor and GitHub Copilot appear only as sync-rules targets, not as skill- or MCP-managed clients like the other three, and no timeline is given for changing that. The package ships as @helter/agentbridge under the MIT licence and exposes a TypeScript SDK with detectInstalledAgents, linkAgentsToHub, syncMcpConfigs, syncProjectRules and runDiagnostics for use outside the CLI.