MCP server

hscli mcp serve turns the whole CLI into a Model Context Protocol server. Point Claude (Desktop, Code, or any MCP client) at it and the agent gets one tool per command — conv_list, docs_article_create, report_conversations — with typed inputs and structured JSON results. No glue code, no wrapper API.

Connect it

The server speaks MCP over stdio: the client launches hscli mcp serve as a subprocess.

Claude Code:

claude mcp add hscli -- hscli mcp serve

Claude Desktop — add to claude_desktop_config.json:

claude_desktop_config.json

{
  "mcpServers": {
    "hscli": { "command": "hscli", "args": ["mcp", "serve"] }
  }
}

The server runs as you — it reuses the OAuth token and Docs API key already in your OS keychain. Run hscli auth login (and hscli docs auth for knowledge-base tools) first.

Reads by default, writes on request

By default only read-only tools are exposed (list, get, search, report, …) — an agent can observe your help desk but not change it. To enable mutating tools (create/update/delete/bulk), start the server with --allow-writes:

claude mcp add hscli -- hscli mcp serve --allow-writes

claude_desktop_config.json

{
  "mcpServers": {
    "hscli": { "command": "hscli", "args": ["mcp", "serve", "--allow-writes"] }
  }
}

--allow-writes is a one-time setup choice — the agent never passes it. When enabled, every tool is available; when not, write tools simply aren’t listed.

Two layers of safety

Write tools are gated by --allow-writes, and the destructive ones — delete, remove, and bulk operations — carry an MCP destructiveHint so the client (e.g. Claude Desktop) prompts you before running them. Reads are marked readOnlyHint; other writes (create/update) are marked readOnlyHint:false + destructiveHint:false, so clients neither auto-approve nor specially warn on them.

How it maps

• One tool per command. conv:bulk-status → conv_bulk_status. The escape-hatch api command and the streaming conv watch are not exposed.
• Flags become typed inputs. Required flags/args are required; --status active|closed becomes an enum; repeatable flags become arrays. Output-shaping flags (--output, --jq, --fields) are dropped — the server always requests JSON.
• Structured results. A command’s JSON is returned as structuredContent (arrays are wrapped under results); plain-text confirmations (“Deleted article 5”) come back as text.

Try it

Once connected, ask the agent things like:

“List the active conversations tagged billing and summarize the top themes.”

With --allow-writes on, the full loop from the home page works — the agent reads the inbox, drafts an article, and publishes it with docs_article_create.

See the command reference for the mcp serve flags.