Agent skill (CLI)
The simplest way to give an agent Siple: a skill that teaches it to run siple-cli. Each action is one command that connects to the running Siple, does one thing and exits. There is no MCP server or other helper process to keep running; only Siple itself runs (it has to, to stay registered and receive calls).
Install
siple-cli skill prints the skill for the version you have installed, so it always matches the binary.
mkdir -p ~/.claude/skills/siple
/Applications/Siple.app/Contents/MacOS/siple-cli skill > ~/.claude/skills/siple/SKILL.md# from the extracted siple-<version> folder
mkdir -p ~/.claude/skills/siple
./bin/siple-cli skill > ~/.claude/skills/siple/SKILL.md# from the folder you unzipped Siple into
New-Item -ItemType Directory -Force "$HOME\.claude\skills\siple" | Out-Null
.\siple-cli.exe skill | Set-Content -Encoding ascii "$HOME\.claude\skills\siple\SKILL.md"For a single project, put it in .claude/skills/siple/SKILL.md in that project instead. Other agents that read Agent Skills use the same file in their own skills folder; agents without skill support can be given its text as instructions.
Reinstall after updating Siple so new methods are covered.
What the skill teaches
- Find
siple-cliand check that Siple is running (siple-cli ctl status). - Call any method with
siple-cli ctl <method> '<json>': the result is JSON on stdout, an error is JSON on stderr with exit code 1. - Look methods up instead of guessing:
siple-cli methodsandsiple-cli methods --jsonwork offline. - Wait without a background process:
call.wait_forandevents.pollwithsinceset to theevent_seqof the action, so nothing is missed between commands. - Common tasks (dial, answer, transfer, DTMF, speak and listen, extensions, settings) and safety rules: confirm before calling or transferring anyone the user did not name, and never print passwords or the control token.
The full text is in the repository at skills/siple/SKILL.md.
Skill or MCP?
| Skill + CLI | MCP (siple-cli mcp) | |
|---|---|---|
| Extra process | none | one stdio server per agent session |
| Tools in the agent's context | one short skill; methods looked up on demand | one tool per method (about 75) |
Agent audio (call.listen audio, call.rx_audio) | WAV as base64 in JSON, or written to a file | native audio content blocks |
| Live event stream | events.poll long-poll | wait_for_event tool |
Both reach exactly the same methods. Use the skill unless your agent cannot run shell commands or you need audio returned as audio content.
