Agent-SIP Ideal for Hermes and n8n workflows

Agent-SIP — a self-hosted SIP voice bridge that can be controlled with AI agents via MCP.

Bring an AI agent to an ordinary phone extension. Agent-SIP registers with your PBX, sends and receives G.711 audio over RTP, and gives you a web dashboard, automation webhooks, and MCP tools for controlling calls from an assistant.

OpenAI's official Realtime SIP integration is designed around Twilio's cloud telephony, but Agent-SIP works with any local SIP server—including Asterisk, FreePBX, 3CX, Kamailio, or a carrier trunk. It speaks SIP directly using UDP signaling and RTP audio, so it doesn't rely on a cloud telephony provider. In our own setup, it runs as extension 500 on a FreePBX box.

Call dashboard

Settings

Agent prompts

Select any screenshot to open the full-size image.

🏗️ Architecture

✨ Features

• Real phone calls through a SIP extension using UDP signaling and RTP audio (PCMU or PCMA).
• OpenAI Realtime speech with natural multilingual conversations, French defaults, configurable language and the marin voice.
• CALL BRIEF objectives that tell the agent what to accomplish on each outbound call.
• MCP control to make calls, steer the agent, speak, hang up, retrieve transcripts, inspect status, and save messages.
• Automation webhooks for call and transcript events—ideal for n8n, Hermes, or your own service.
• Background office ambience with selectable bundled sounds and adjustable volume.
• Authenticated web UI for calls, live transcripts, logs, configuration, and prompts.
• Simple deployment with Docker and a published GHCR image.
• Call safeguards including configurable ring limits, ring timeouts, maximum agent turns, and automatic hangup after goodbyes.

🚀 Quick start with Docker

Agent-SIP works best with Docker host networking because SIP/SDP embeds network addresses and RTP uses a UDP port range.

make setup # creates .env from the template (first time)# edit .env: VOICE_API_KEY + your PBX details
make up # builds and starts

Without Make, use the published image directly. Edit .env and, at minimum, provide your SIP server, extension credentials, reachable SIP_ADVERTISE_HOST, and VOICE_API_KEY:

docker pull ghcr.io/ai-redcode/agent-sip:latest
cp .env.docker.example .env
docker run -d --name agent-sip --network host --env-file .env \
-v ./var:/app/var \
ghcr.io/ai-redcode/agent-sip:latest

Open http://localhost:8090 and sign in with admin / admin.

Warning

Change the default web password immediately, especially before exposing the UI beyond a trusted local network.

Host networking is required for the normal Docker setup: the PBX must be able to reach the SIP address and RTP ports advertised inside SDP. By default Agent-SIP uses TCP 8090 for the UI, TCP 8765 for MCP, UDP 5062 for SIP, and UDP 40000–40100 for RTP.

The Makefile prefers Docker Compose when available and falls back to the docker run command above when it is not.

🖥️ Web UI

The dashboard is organized into three tabs:

Tab	What it does
Call	Place and end calls, enter a CALL BRIEF, inject speech, follow the live transcript, and inspect recent logs.
Settings	Configure the SIP endpoint, voice provider and speaking speed, background noise, MCP, webhooks, and UI credentials.
Agent	Set the agent name, caller ID, default language, inbound context, inbound/outbound prompts, ring limits, and automatic-hangup behavior.

The status bar keeps the essentials visible at a glance: SIP Registration, Voice Provider, Call State, Active Call, and Agent. Settings are grouped into focused boxes for SIP endpoint, Voice provider (including speaking speed), Background noise, MCP, Webhook, and UI Password. Saving configuration persists it to var/config.json; API responses mask stored secrets.

🔧 MCP usage

The MCP control API listens at http://127.0.0.1:8765 by default. Set MCP_AUTH_TOKEN and send it as a Bearer token. Standard MCP clients can launch the included agent-sip-mcp stdio bridge, which proxies tools to the running HTTP service.

Tool	Arguments	Purpose
get_status	—	Return SIP registration, current call state, and recent call details.
make_call	number, call_brief (optional)	Start an outbound call with a per-call objective.
hangup_call	—	End the active call.
say	text	Speak text into the active call.
steer	instructions and/or speed	Change instructions, tone, or speaking speed (0.25–4.0) mid-call.
get_transcript	—	Return recent transcript messages.
save_message	recipient, caller_name, message, callback_number, language, confirmed_by_caller	Save a caller-confirmed message.

Call a tool directly over HTTP:

curl -X POST http://localhost:8765/call \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_MCP_TOKEN" \
-d '{"name":"make_call","arguments":{"number":"201","call_brief":"Bonjour, ..."}}'

What is a CALL BRIEF?

A CALL BRIEF is objective text injected into the Realtime agent's session instructions for that call—not merely an opening sentence. It defines the task throughout the conversation: who to call, what to ask, what may be disclosed, and what result to collect. It can also set the language or tone, for example: Speak in Armenian, introduce yourself warmly, and ask whether Tuesday at 14:00 is available.

📡 Webhooks

Ideal for n8n workflows: configure a primary WEBHOOK_URL and, optionally, WEBHOOK_URL2. The second destination is useful when the same events should also flow to an n8n workflow. Delivery is best effort and never blocks call signaling or audio.

Event	When it is sent
call.started	A call begins.
transcript.partial	Partial speech is available, if enabled for that call direction.
transcript.final	A finalized transcript item is available.
call.ended	The call ends; includes outcome, rings, and duration.

Every JSON payload includes event and type (with the same value), plus call_id, caller_number, called_number, agent_name, transcript, timestamp, and mcp_url. A typical ended-call payload looks like this:

{
"event": "call.ended",
"type": "call.ended",
"call_id": "abc123",
"caller_number": "200",
"called_number": "201",
"agent_name": "Reception",
"transcript": [{"role": "agent", "text": "Bonjour."}],
"outcome": "completed",
"rings": 2,
"duration": 47.3,
"timestamp": "2026-08-05T12:00:47+00:00",
"mcp_url": "http://127.0.0.1:8765"
}

To replay a representative event against your receiver while developing:

curl -X POST http://localhost:5678/webhook/agent-sip \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_WEBHOOK_TOKEN" \
-d '{"event":"call.started","type":"call.started","call_id":"test-1","caller_number":"200","called_number":"201","agent_name":"Agent","transcript":[],"timestamp":"2026-08-05T12:00:00+00:00","mcp_url":"http://localhost:8765"}'

When WEBHOOK_AUTH_TOKEN is set, Agent-SIP sends both Authorization: Bearer … and X-Hub-Signature-256: sha256=…. The HMAC-SHA256 signature is calculated over the exact JSON body using the same token as the secret. (X-Hub-Signature with HMAC-SHA1 is also provided for compatibility.) Use WEBHOOK_NOTIFY_PARTIALS_INCOMING and WEBHOOK_NOTIFY_PARTIALS_OUTGOING to control noisy partial events independently.

n8n integration: create a Webhook node, place its production URL in WEBHOOK_URL2, verify the signature in the first workflow step, and route on the type field. A call.ended branch can summarize the transcript, update a CRM, or notify a home channel.

🤖 Hermes Agent integration (PersonalAssistant)

Ideal for Hermes workflows: Hermes can use Agent-SIP as both a callable tool server and an event source.

1. Create an executable bridge wrapper named agent-sip-mcp-bridge. For a remote Agent-SIP host, its core command can be:

   #!/usr/bin/env bashexec ssh voice-host "MCP_AUTH_TOKEN=YOUR_MCP_TOKEN /tmp/agent-sip/.venv/bin/agent-sip-mcp"

   ssh does not automatically forward locally exported environment variables. Pass MCP_AUTH_TOKEN inline in the remote command as shown. For a non-default control URL on the remote host, pass AGENT_SIP_API_URL=… inline too.

2. Register the stdio bridge with Hermes:

   hermes mcp add agent-sip --command /path/to/agent-sip-mcp-bridge

3. In the Hermes webhook platform, subscribe to call.started, transcript.final, and call.ended (and transcript.partial only if needed). Use the same HMAC secret as WEBHOOK_AUTH_TOKEN, then route deliveries to your home channel.

4. Ask your assistant: “Call X and ask about Y.” Hermes builds the CALL BRIEF, invokes make_call, follows the resulting event/transcript flow, and summarizes the outcome.

🧪 Development

Requires Python 3.11 or newer:

python3 -m venv .venv
.venv/bin/pip install -e .
.venv/bin/agent-sip

Run the test suite:

.venv/bin/python -m pytest tests/ -v

Repository layout:

app/ SIP, RTP, Realtime, MCP, webhook, and web application code
assets/ambient/ Bundled background sound loops
docs/screenshots/ Web UI screenshots used in this README
scripts/ Diagnostic utilities
static/ Dashboard and login pages
tests/ Unit and integration tests
tools/ Development helpers
var/ Persisted runtime configuration and messages
Dockerfile Container image definition
docker-compose.yml Host-networked deployment

The FastAPI documentation is available at http://localhost:8090/docs while the service is running.

⚙️ Configuration reference

Environment variables override settings that are not already populated in persisted var/config.json. VOICE_API_KEY is required for live speech.

SIP

Variable	Default	Description
SIP_SERVER_HOST	127.0.0.1	FreePBX/Asterisk host.
SIP_SERVER_PORT	5060	PBX SIP port.
SIP_TRANSPORT	udp	SIP transport; only UDP is supported.
SIP_USERNAME	200	SIP extension/username.
SIP_AUTH_USERNAME	empty	Authentication username; falls back to the extension where applicable.
SIP_PASSWORD	empty	SIP password.
SIP_LOCAL_HOST	0.0.0.0	Local bind address.
SIP_LOCAL_PORT	5062	Local SIP UDP port.
SIP_ADVERTISE_HOST	127.0.0.1	Address advertised to the PBX in SIP/SDP.
SIP_RTP_PORT_START / SIP_RTP_PORT_END	40000 / 40100	RTP UDP port range.
SIP_CODEC	pcmu	G.711 codec: pcmu or pcma.

Voice and agent

Variable	Default	Description
VOICE_PROVIDER	openai	Voice provider; currently OpenAI only.
VOICE_API_KEY	required	OpenAI API key.
VOICE_BASE_URL	wss://api.openai.com/v1/realtime	Realtime WebSocket endpoint.
VOICE_MODEL	gpt-realtime-2.1	Realtime model name.
VOICE_VOICE	marin	OpenAI Realtime voice.
VOICE_SPEED	1.0	Speaking speed (0.25–4.0).
AGENT_NAME	Agent	Displayed agent name.
AGENT_DEFAULT_LANGUAGE	fr	Session language, such as fr, en, or hy.
AGENT_INBOUND_PROMPT / AGENT_OUTBOUND_PROMPT	built in	Direction-specific system prompts.
AGENT_INBOUND_BRIEF	empty	Recipients or context for incoming calls.
AGENT_CALLER_ID	200	Outbound caller identity.
AGENT_MAX_RINGS	6	Maximum 180 Ringing responses (1–20).
AGENT_MAX_RING_SECONDS	30	Outbound ring timeout (5–300 seconds).
AGENT_MAX_AGENT_TURNS	3	Maximum agent turns used by silence handling (1–20).
AGENT_END_GRACE_SECONDS	4.0	Grace period before auto-hangup after a goodbye.

MCP, webhooks, ambience, and UI

Variable	Default	Description
MCP_ENABLED	true	Compatibility setting; the MCP service is always enabled.
MCP_HOST	127.0.0.1	MCP HTTP bind address (.env.docker.example uses 0.0.0.0).
MCP_PORT	8765	MCP HTTP port.
MCP_AUTH_TOKEN	empty	Optional Bearer token.
WEBHOOK_ENABLED	true	Compatibility setting; delivery occurs when a URL is configured.
WEBHOOK_URL / WEBHOOK_URL2	empty	Primary and secondary receiver URLs.
WEBHOOK_AUTH_TOKEN	empty	Bearer token and HMAC secret.
WEBHOOK_NOTIFY_PARTIALS_INCOMING	false	Send partial transcript events for inbound calls.
WEBHOOK_NOTIFY_PARTIALS_OUTGOING	true	Send partial transcript events for outbound calls.
AMBIENT_ENABLED	false	Mix background audio into calls (.env.docker.example enables it).
AMBIENT_FILE	office.wav	File from assets/ambient/.
AMBIENT_VOLUME	0.12	Mix level (0.0–0.5).
WEB_ENABLED	true	Enable the web application.
WEB_USERNAME	admin	Web UI username.
WEB_PASSWORD	admin	Web UI password—change it.

License and disclaimer

No license file is currently included in this repository; all rights remain with the copyright holder unless a license is added. Agent-SIP can place real telephone calls—follow local calling, recording, consent, privacy, and emergency-services laws, and secure all credentials before deployment.