CLI + MCP

Your data, inside any agent

Calliope Data speaks CLI and MCP. Plug your business data into Claude, ChatGPT, Gemini, Cursor or your own harness — natural-language answers with cited sources, and business rules your agent cannot bypass.

View on GitHub Request Demo

brew install Calliope-AI/tap/calliope
calliope auth login --api-key <key>
calliope orgs use acme
calliope ask "how are sales this quarter?"

Or skip the install entirely — add this MCP server to your agent: https://mcp.calliope.so/mcp

Works with the agents you already use

Anything that speaks MCP over HTTP, or that can run a binary on your PATH.

Claude Code
Claude Desktop
ChatGPT
Gemini
Cursor
ElevenLabs
Your own harness

Command line

Calliope in your terminal

Twenty-two commands: sixteen in six groups — auth, orgs, config, docs, concepts, rules — plus ask, query, schema, doctor, skill and version. Every command returns the same envelope, and the breadcrumbs tell you what to run next.

Readable in a terminal, JSON in a pipe — automatically
--json, --quiet, --md, --jq '<expr>' — the filter runs inside the binary, so no jq needed — and --csv on calliope query
Stable exit codes: 0 ok · 1 error · 2 usage · 3 unauthorized · 4 not found · 5 limit exceeded
Layered config — flags, CALLIOPE_* env, project and global files. calliope config list shows which layer each value came from

{"ok": true, "data": [...], "summary": "12 documents",
 "breadcrumbs": [{"action": "show",
                  "cmd": "calliope docs show <id>"}]}

Coding agents

The skill ships inside the binary

calliope skill prints the full SKILL.md — invariants, command catalogue and recipes — exactly as the installed version packaged it. A test compares its command catalogue against the real command tree and fails the build if the two drift, so it cannot fall out of step with the binary your agent is actually calling.

A Claude Code plugin that installs the skill and a /calliope command
It warns at session start when the binary or the credential is missing
Domain skills are plain Markdown composing commands — no Go required
calliope doctor diagnoses install, credential and connectivity

calliope skill

Model Context Protocol

Calliope as a remote MCP server

Point your agent at the hosted server. Two ways in, both live today.

Server URL: https://mcp.calliope.so/mcp

OAuth 2.1

Clients like Claude, ChatGPT and Gemini: paste the URL, sign in, pick your organization. Dynamic client registration and PKCE S256. The organization comes from the token, never from a header.

API key

Servers and voice agents such as ElevenLabs send their own key. Calliope stores no credentials — it only relays them. Create a key in the Calliope Data UI, under API keys.

X-Calliope-Api-Key: <your key>
X-Calliope-Org:     <org name>

Tool	What it does
ask_calliope	Natural-language questions over data and documentation. Returns an answer, cited sources and aggregated data. No SQL.
search_documentation	Semantic search across documentation, with document, page and section for every fragment.
list_documents	Lists available documents, filterable by text or tag.
get_document	Document metadata by id.
list_concepts	Business concepts in the ontology — what data exists.
describe_concept	Description and attributes of one concept.
list_business_rules	The organization's shared business rules.

Streamable HTTP for remote clients, stdio for local development
Three explicit scopes: calliope:catalog:read, calliope:data:read, calliope:query:execute
orgName is never a tool argument — it is resolved server-side from the request credentials

Governance

Your MCP agents never write SQL

Access without a blank cheque. Over MCP, the seven tools take questions, not queries — an agent cannot reach past them. The CLI keeps query and schema for people who want raw SQL, and its own skill tells agents to reach for ask first.

No SQL over MCP

The MCP tools take natural language and return answers with cited sources. No tool accepts a query, so an agent connected this way cannot write one.

Rules enforced at query time

Business rules are first-class objects applied when the query runs — not suggestions in a prompt. Resistant to prompt injection and conversational drift.

Isolated per organization

The tenant is resolved on the server, from the credential — never from an argument the agent controls.

Every answer cites its sources

Document, page and section, returned alongside the answer.