A gateway that connects any A2A (Agent-to-Agent) protocol agent to Telegram, Slack, WhatsApp, Google Chat, Email, Discord, and your own custom channels. You build the agent, the gateway handles the channels.

 Telegram ──┐                        ┌─────────────────────┐
    Slack ──┤                        │   Your A2A Agent    │
 WhatsApp ──┤                        │   (ADK, LangGraph,  │
   G Chat ──┼──▶  a2a-gateway  ──A2A─┤    CrewAI, custom…) │
    Email ──┤  ◀──── /push (A2A) ────┤                     │
  Discord ──┤                        └─────────────────────┘
   Custom ──┘

Adding a built-in channel is mostly just setting env vars. Need a platform we don't support? Subclass SimpleChannel, implement one method, and you're in. The gateway takes care of message chunking, rate limiting, retries, debouncing, typing indicators, streaming responses, and per-channel markdown formatting. Everything is opt-in: if you don't configure a feature, it doesn't run.

docker run -p 8000:8000 \
  -e A2A_SERVER_URL=http://host.docker.internal:8001 \
  -e TELEGRAM_BOT_TOKEN=your-bot-token \
  ghcr.io/eliasecchig/a2a-gateway:main

Get a bot token from @BotFather in under a minute.

Prerequisites: Python 3.12+ and uv.

git clone https://github.com/eliasecchig/a2a-gateway.git
cd a2a-gateway
uv sync

Run with env vars (no config file needed):

export A2A_SERVER_URL=http://localhost:8001
export TELEGRAM_BOT_TOKEN=your-bot-token
uv run a2a-gateway

Or with a config file:

cp config.example.yaml config.yaml
uv run a2a-gateway --config config.yaml

Or with the built-in test agent (a minimal A2A echo server — no LLM, no credentials needed):

uv run a2a-gateway --with-agent

The gateway starts on http://localhost:8000.

By default the gateway is single-user: one bot, one A2A agent, zero infra beyond the TELEGRAM_BOT_TOKEN env var. The simple case above stays simple.

For multitenant deployments — one bot on any channel (Telegram, Slack, WhatsApp, Discord, Google Chat, Email, custom) serving N users, each linked to a distinct A2A user identity, with proactive notifications via /push targeting a2a_user_id instead of raw chat IDs — opt in with MULTITENANT_MODE=true and set GATEWAY_ADMIN_TOKEN. Users enroll via https://t.me/...?start=<token> on Telegram or by typing connect <token> in DM on any other channel. State lives in SQLite by default (zero infra) or any asyncpg-compatible Postgres in production. See docs/multitenancy.md.

All channels use official APIs.

Every inbound message goes through this pipeline. Each step is optional and controlled by config.

Inbound message
  │
  ├─ ACK reaction (emoji / read receipt)
  ├─ Debounce (coalesce rapid messages)
  ├─ Group policy check (open / mention-only / disabled)
  ├─ Rate limit (A2A side)
  ├─ Concurrency limit
  ├─ Typing indicator
  │
  ├─ ──▶  A2A agent  ──▶  response (or SSE stream)
  │
  ├─ Markdown adaptation
  ├─ Chunking (split long responses, preserve code fences)
  ├─ Rate limit (channel side)
  │
  └─ Send reply (or stream edits in real time)

A prebuilt image is published to GHCR on every push to main (see Quick start for usage). To build locally:

docker build -t a2a-gateway .
docker run -p 8000:8000 \
  -e A2A_SERVER_URL=http://your-agent:8001 \
  -e TELEGRAM_BOT_TOKEN=your-bot-token \
  a2a-gateway

Mount a config file: -v $(pwd)/config.yaml:/app/config.yaml.

uv run pytest tests/ -v                  # all offline
uv run pytest -m live tests/live/ -v     # live integration tests

See CONTRIBUTING.md.

Copyright 2026 Google LLC. Licensed under the Apache License, Version 2.0. See LICENSE.