incant-mcp

MCP server for Incant — lets AI agents author, test, review, publish, and target prompts, doing everything the console can (minus deployment administration) within the configured API key's roles.

Setup

{
  "mcpServers": {
    "incant": {
      "command": "uvx",
      "args": ["incant-mcp"],
      "env": {
        "INCANT_URL": "https://prompts.internal",
        "INCANT_API_KEY": "incant_sk_...",
        "INCANT_ENVIRONMENT": "staging"
      }
    }
  }
}

INCANT_ENVIRONMENT is optional (tools default to the deployment's default environment and accept a per-call override). Add --read-only to the args to register only the read/test tools.

Permissions

The server adds nothing and hides nothing: every call runs under the API key's roles, enforced by the deployment. A viewer key can explore and render-test but not publish; a 403 surfaces the server's role explanation verbatim. Protected environments additionally demand a confirm echo on mutations — the paired skills instruct agents to relay that confirmation to a human first.

Tools

• Discover: list_prompts, get_prompt, list_rules, list_environments, get_publish_history, get_targeting_history, get_audit
• Test: render_prompt (real serving path, with pin replay), evaluate_targeting, diff_versions
• Author: create_prompt, edit_draft (create/update/get/render/diff/discard), commit_draft, review_draft, set_prompt_metadata
• Release & target: publish_prompt (pointer + optional default promotion + rule archive, atomic), rollback_pointer, upsert_rule, set_rule_status, upsert_segment, set_default, kill_switch, rollback_targeting

Skills

The skills/ directory pairs this server with agent skills — incant-authoring (draft → test → review → commit) and incant-release (publish → target → roll out → roll back) — that encode the workflows and guardrails. Install by copying into your agent's skills directory (e.g. .claude/skills/).