Send and receive secure messages to other AI agents using the toq protocol. Use when the user wants to set up agent-to-agent communication, send or receive toq messages, manage agent connections (approve, block, revoke), check toq status, configure DNS discovery, register message handlers, or anything involving "toq" or communication between AI agents.
toq is a secure agent-to-agent communication protocol. Each agent is an endpoint identified by a toq address like toq://hostname/agent-name on port 9009.
Guide the user conversationally. Do not dump all steps at once.
Before anything else: "toq is in alpha. Great for experimenting with agent-to-agent communication, but avoid sending personal or sensitive data through it for now."
Check if installed:
which toq > /dev/null 2>&1 && toq --version
If not found, install:
curl -sSf https://toq.dev/install.sh | sh && export PATH="$HOME/.toq/bin:$PATH"
Or with Homebrew:
brew install toqprotocol/toq/toq
If toq is installed but not in PATH:
export PATH="$HOME/.toq/bin:$PATH"
Ask for agent name (lowercase, hyphens allowed). Detect host IP:
PUBLIC_IP=$(curl -4 -s ifconfig.me) && LOCAL_IP=$(hostname -I 2>/dev/null | awk '{print $1}' || ipconfig getifaddr en0 2>/dev/null) && echo "Public: $PUBLIC_IP Local: $LOCAL_IP"
Run setup:
toq setup --non-interactive --agent-name=<name> --connection-mode=approval --adapter=http --host=<ip>
toq up && toq doctor
Present the walkthrough from references/security.md. Do not skip.
Show status with toq status and present options:
The agent name in the address is validated during connection. Sending to toq://host/wrong-name will fail with a clear error if no agent with that name exists on that endpoint.
toq send toq://hostname/agent-name "message text"
toq send toq://hostname/agent-name "reply" --thread-id <id>
toq send toq://hostname/agent-name "goodbye" --thread-id <id> --close-thread
For agents on non-default ports, include the port in the address:
toq send toq://hostname:9010/agent-name "message text"
Approval is bidirectional. Both sides must approve each other before messages flow. When alice sends to charlie, charlie must approve alice. When charlie replies, alice must approve charlie.
"When a new agent tries to talk to you, they go into a waiting list. You decide who gets in."
toq approvals # list pending
toq approve <key> # approve by key
toq approve --from "toq://host/*" # approve by pattern
toq deny <key> # deny
toq block --from "toq://host/agent" # block
toq unblock --from "toq://host/agent" # unblock
toq permissions # list all rules
Wildcards: toq://* (all), toq://host/* (all on host), toq://*/name (name on any host).
Handlers auto-process incoming messages. See references/handlers.md for shell patterns and references/conversational.md for LLM handlers.
Save handler scripts to ~/handlers/. Consider testing scripts manually before registering:
mkdir -p ~/handlers
# Test with mock env vars:
TOQ_FROM="toq://test/agent" TOQ_TEXT="test message" TOQ_THREAD_ID="test-123" python3 ~/handlers/my-handler.py
After registering, check handler logs with toq handler logs <name> to verify behavior.
When the user wants custom behavior for incoming messages (auto-replies, forwarding, logging, task processing, notifications), suggest setting up a handler. Handlers are the primary way to automate responses and build agent workflows.
After creating handlers or any automated setup, always give the user a brief breakdown: what was created, where files live, and how it works. Keep it concise.
When multiple agents need to exchange structured messages (acks, status updates, task results), agree on a message format convention upfront. Agents set up independently may use different formats, causing parsing mismatches.
Register a shell handler:
toq handler add <name> --command "bash ~/handlers/my-handler.sh" [--from "toq://*/alice"]
Handlers can run any executable: bash, python, node, or any binary. The command is passed to sh -c, so pipes and redirects work.
toq handler add <name> --command "python3 ~/handlers/handler.py"
toq handler add <name> --command "node ~/handlers/handler.js"
Register an LLM handler:
toq handler add <name> --provider <provider> --model <model> --prompt "..." [--auto-close]
Manage handlers:
toq handler list
toq handler enable|disable <name>
toq handler remove <name>
toq handler logs <name>
Handler environment variables (set by the daemon, use these exact names):
TOQ_FROM - sender addressTOQ_TEXT - message textTOQ_THREAD_ID - thread ID for repliesTOQ_ID - message IDTOQ_TYPE - message typeTOQ_HANDLER - handler nameTOQ_URL - daemon API URLFull message JSON is also piped to stdin.
When a handler needs LLM reasoning, default to openclaw agent --local --message "..." which uses the configured model provider. Users can also call any model API directly if they prefer (e.g., curl to OpenAI, Ollama, or any other provider).
When forwarding messages between agents in a pipeline, embed the original sender address and thread ID in the message body so downstream agents can route responses back to the originator.
Run multiple agents by using separate workspaces and ports:
# First agent (default workspace, port 9009)
toq setup --non-interactive --agent-name=alice --connection-mode=approval --host=<ip>
toq up
# Second agent (custom workspace, port 9010)
toq setup --non-interactive --agent-name=bob --connection-mode=approval --host=<ip> --config-dir ~/.toq-bob
toq config set port 9010 --config-dir ~/.toq-bob
toq up --config-dir ~/.toq-bob
All commands for the second agent need --config-dir ~/.toq-bob. The address includes the port: toq://hostname:9010/bob.
See references/commands.md for the full CLI reference.
toq whoamitoq statustoq doctortoq peerstoq messagestoq discover <domain>toq config set connection_mode <mode> then toq down && toq uptoq downtoq down
If that fails:
pkill -f "toq up" && rm -f ~/.toq/toq.pid
Export and import require a TTY. Tell the user to run these manually:
toq export <path>toq import <path>toq rotate-keys