What's new
Changes to the Chatley MCP server — new tools, changed behaviour, and anything that affects a connection you have already made.
Changes to the MCP server, newest first. Anything that could change how an existing connection behaves is marked, because a tool that quietly starts answering differently is worse than one that visibly stops working.
Tool names and error codes on this page are stable. When one has to change, it appears here first and the old form keeps working for a stated period.
First release
The initial tool set. Sixteen tools, all read or safe-write.
Calls
list_calls, get_call_details, get_call_transcript, get_call_analytics.
Transcripts are a separate tool from the listing on purpose: a listing that carried them would fill your assistant's context after a handful of calls and make everything after it worse. List first, then fetch the one you want.
Call outcomes use a fixed set of values — caller_ended, agent_ended,
transferred, no_answer, voicemail, busy, abandoned,
max_duration_reached, blocked, failed, in_progress, unknown — so you
can branch on them safely.
AI voice agents, with drafts
list_agents, get_agent, create_agent, update_agent,
publish_agent_draft.
The behaviour to know about: update_agent saves a draft. Your agent
keeps answering calls exactly as before until publish_agent_draft is called.
This is deliberate — an assistant acting on a misread instruction should not be
able to change what a caller hears in one step.
get_agent returns any unpublished draft alongside the live configuration, and
tells you which fields differ. If a change seems not to have taken effect, look
there first: an unpublished draft is the usual reason.
Saving again replaces the draft rather than adding to it, so put everything you want staged into one call.
Voices
list_voices. Voice IDs are case-sensitive — pass them back exactly as
returned. A voice's ID and its name are not always the same word, so use the
name when talking to a person and the ID when calling a tool.
Knowledge bases
list_knowledge_bases, delete_knowledge_base_items.
Deleting takes two calls. The first returns exactly what would be removed,
including how many agents currently use each one, and changes nothing. Only a
second call carrying confirmed: true removes anything, and that check is ours
— a client cannot skip it.
Account
get_workspace_info, list_tenants.
If you administer more than one Chatley account, list_tenants tells you which,
and every other tool accepts a tenant_id to target one. Naming an account you
do not administer is refused and recorded, so there is no point guessing.
Metered access
get_rate_card, get_balance.
Reading is free — calls, transcripts, analytics, agents, voices and knowledge bases are never metered, so an assistant can explore your account without spending anything. At zero balance metered operations stop and reading carries on.
If get_rate_card reports prices_are_provisional: true, the numbers are not
final; treat them as indicative.
Known limitations
Stated here rather than discovered later.
No contacts or campaigns. Those tools arrive with the campaign engine.
No adding to a knowledge base. A Chatley knowledge base is an uploaded document, and an upload is not something a tool call can carry. Add it in Chatley; the agent picks it up.
Deleting an AI voice agent, deleting a contact, and bulk updates are not offered at all. These are the operations where one confused call does the most damage. Do them in Chatley, where the consequences are visible on screen.
One region. A single global endpoint. If data residency matters to you, tell us — regional endpoints exist as a plan, not as a deployment.