Cross-vendor agent-to-agent protocol — Claude, Codex, and Gemini communicate via file-based P2P messaging.
tap is a growing TypeScript project in the AI payments / x402 ecosystem, focused on a2a, agent-to-agent, ai-agents, claude. It currently has 20 GitHub stars and 5 forks, and sits alongside related tools like suede-creator-skills, nexus, npm:wispy-ai, Antigravity-Super-Skill-Architecture, capiscio-mcp-python, portable-agent-usb.
@hua-labs/tap is a lightweight CLI that lets multiple AI agent runtimes, such
as Claude and Codex, exchange messages through a shared file-backed
communication directory.
It focuses on safe setup, durable delivery evidence, and runtime-specific diagnostics without requiring a central daemon for every runtime.
0.6.x is an advanced operator preview.
agent-a and agent-b.setup is dry-run-first and only applies reviewed tap-managed changes.0.6.x.For detailed AI-operator troubleshooting, concepts, and advanced runtime notes, read AI_GUIDE.md. AI operators should read it before claiming live delivery or runtime readiness.
>=22.6.0Run inside an existing git repository. For a new test directory only:
git init
Then inspect and apply setup:
# 1. Inspect setup changes first
npx @hua-labs/tap setup --profile codex-cli --dry-run --json
# 2. Apply only reviewed setup-safe changes
npx @hua-labs/tap setup --profile codex-cli --apply --json
# 3. Check setup readiness
npx @hua-labs/tap doctor --setup --profile codex-cli --json
npx @hua-labs/tap status --json
To add and verify a concrete runtime lane later:
npx @hua-labs/tap add codex --name agent-a
npx @hua-labs/tap ready --surface codex-cli --agent agent-a --apply --json
npx @hua-labs/tap comms-doctor --all-known --json
tap setup is dry-run by default. With --apply, the reviewed public setup
path creates tap-owned directories, an initial tap state file, and guarded
tap-managed repo .mcp.json entries.
It does not start receiver, projection, uplink, bridge, app-server, headless runner, or remote panel processes. It does not publish presence, repair route tuples, send messages, or read credentials.
Most users should start with setup. Use init only when you want to manually
create or reset the shared communication directory and local tap state.
| Profile | Use when |
|---|---|
codex-cli |
Codex CLI or headless CLI will use MCP tools plus inbox/receiver paths. |
codex-app |
Codex App/Desktop route readiness should be inspected read-only. |
claude-channel |
Claude/channel readiness should be inspected read-only. |
setupInspect and optionally apply first-run public deployment changes.
npx @hua-labs/tap setup --profile codex-cli --dry-run --json
npx @hua-labs/tap setup --profile codex-cli --apply --json
npx @hua-labs/tap setup --profile codex-cli --profile-pack ./tap-profile-pack.json --json
tap refuses to mutate user-managed or ambiguous MCP entries.
readyPrepare or verify a concrete agent on a runtime surface.
npx @hua-labs/tap ready --surface codex-cli --agent agent-a --json
npx @hua-labs/tap ready --surface codex-cli --agent agent-a --apply --json
By default, ready reports the current or planned readiness state. With
--apply, it applies reviewed readiness changes for the selected surface and
agent.
initCreate or reset the shared comms directory and .tap-comms/ state.
npx @hua-labs/tap init
npx @hua-labs/tap init --comms-dir /path/to/comms
npx @hua-labs/tap init --permissions safe
npx @hua-labs/tap init --permissions full
add <runtime>Add a runtime instance and patch the runtime config when verification allows it.
npx @hua-labs/tap add claude
npx @hua-labs/tap add codex --name agent-a
npx @hua-labs/tap add codex --name agent-b --port 4520
gemini remains accepted for legacy compatibility, but new first-run docs do
not recommend it as a default runtime.
statusShow installed runtimes and their status.
npx @hua-labs/tap status
npx @hua-labs/tap status --json
doctorDiagnose setup, config drift, bridge health, managed MCP wiring, and runtime state.
npx @hua-labs/tap doctor
npx @hua-labs/tap doctor --setup --profile codex-cli --json
comms-doctorExplain delivery by runtime surface, including local inbox evidence and live adapter readiness.
npx @hua-labs/tap comms-doctor --all-known --json
npx @hua-labs/tap comms-doctor --agent agent-a --plan-send --json
flow-doctorDiagnose one receiver/promoter lane without mutating process state.
npx @hua-labs/tap flow-doctor --agent agent-a --json
flow-doctor is read-only by default. Its reviewed cleanup mode only archives
stale non-lane presence records with a manifest and prunes matching stale
heartbeat entries.
reviews registerRegister formal review outcomes into a discoverable review evidence stream.
npx @hua-labs/tap reviews register --source ./tap-comms --dry-run --json
serveStart the bundled tap MCP server on stdio.
npx @hua-labs/tap serve
bridge start|stop|restartManage Codex app-server bridge lifecycle.
npx @hua-labs/tap bridge start codex --agent-name agent-c
npx @hua-labs/tap bridge stop codex --keep-server
npx @hua-labs/tap bridge restart codex
agent-a, that sends or receives
messages.codex-cli or
codex-app.Safe mode reduces destructive local file operations. It is not a network-isolated mode.
safe is the default.full enables less restricted local runtime settings and should only be used
on trusted machines.See AI_GUIDE.md for details.
Most operational commands support --json for machine-readable output.
npx @hua-labs/tap status --json
Exit codes use 0 for ok and 1 for error. Error codes use the TAP_*
prefix.
| Runtime | Status | Notes |
|---|---|---|
| Claude | supported | Uses .mcp.json / channel-style delivery. |
| Codex | supported | Uses CLI, MCP, and app-server surfaces. |
| Gemini | legacy / deprecated | Kept for compatibility only. |
| Other | custom profile-pack input | Use data-only profile packs in 0.6.x. |
Antigravity CLI is not a bundled adapter in 0.6.x. Model it as a custom
profile-pack surface until a dedicated integration is released.
examples/tap-profile-pack.example.jsonexamples/The narrative examples are real collaboration stories, not setup defaults.
See CHANGELOG.md for full release history.
Highlights in 0.6.x:
ready, comms-doctor, and flow-doctor for surface-specific diagnostics;tap reviews register evidence registration;MIT
25 open-source Agent Skills for Claude Code and Codex: multi-agent orchestration, A-F code review, AI evals, CI ship-gates, design, copy, SEO, iOS shipping, music rights, and consumer recovery (the Amazon negotiator that recovered $448.31).
The AI-to-AI Protocol Layer that catches unreliable agent outputs before they cost you money. Claim-level verification, escrow, and slashing. 224 tests, 12/12 adversarial patterns caught.
Autonomous AI agent with dual engine (Gemini + Claude), Memory Bank, sessions, skills.sh registry, cross-model delegation, 42+ integrations, and x402 agentic commerce
A Federated Multi-Agent System (MAS) architecture for Google Antigravity. Orchestrates specialized AI Skills (Backend, DevOps, Frontend) using the Google ADK and A2A Protocol.
CapiscIO MCP Guard - Secure your MCP tools with agent-to-server authentication, authorization, and audit logging
Run Claude Code and OpenAI Codex directly from a USB drive without installation, leaving no trace on the host machine.
The agent-native LLM router for autonomous agents. 55+ models (8 free), <1ms local routing, USDC payments on Base & Solana via x402.
The living ecosystem where AI agents complete tasks through workflow loops, improve through iterative execution, are evaluated by mentor agents or humans in the loop, and turn completed work into reusable work experience and data to improve future agents.
The living ecosystem where AI agents complete tasks through workflow loops, improve through iterative execution, are evaluated by mentor agents or humans in the loop, and turn completed work into reusable work experience and data to improve future agents.
The trust layer for agent-to-agent commerce — natural-language mandates, ERC-7710 delegated permissions, x402 payments, escrow, and dispute resolution as one open, catch-all Agent Skill / Claude Code plugin.
Open Source AI trading agent that operates autonomously across 1000+ markets - Polymarket, Kalshi, Binance, Hyperliquid, Solana DEXs, 5 EVM chains. Scans for edge, executes instantly, manages risk while you sleep. Agent commerce protocol for machine-to-machine payments. Self-hosted. Built on Claude.
Live data for AI agents — search, research, markets, crypto, X/Twitter. Pay-per-call via x402 micropayments.