#AI agents, API & MCP

AI calling agents make calls at a volume no human team can match, which is exactly what A/B testing needs. Clarity gives your agent the same loop your reps use: get a script, make the call, log what happened.

#Create a key

Admins go to Settings → API, name a key and create it. Copy it right away; it's shown once. Revoke it any time.

#Connect over MCP

Add Clarity to any MCP-compatible agent:

{
  "mcpServers": {
    "clarity": {
      "type": "http",
      "url": "https://claritysdr.com/api/mcp",
      "headers": { "Authorization": "Bearer clr_live_…" }
    }
  }
}

#Tools

Tool	When to use it
`list_experiments`	Find running A/B tests and their variants
`assign_variant`	Before a call: which script this agent should use (balanced per agent)
`log_outcome`	After a call: outcome, optional objection, duration and transcript
`get_results`	Per-variant rates, the verdict at 95% confidence, dials still needed
`get_playbook`	Before a shift: the script to run, behaviors that book more or fewer meetings, likely objections with answers
`analyze_call`	Behavior breakdown of a transcript, stores nothing

Every tool is also available as POST /api/v1/<tool> with a JSON body and the same Authorization header.

#Behaviors

When log_outcome includes a transcript, Clarity measures talk share, questions, the longest monologue, filler words and the call-flow checks. Those measurements feed Behaviors that book meetings on your dashboard. Transcript text is never stored.

Agent calls are tagged with their agent_ref and kept off your reps' leaderboards.

#A self-improving loop

Call get_playbook at the start of every shift (or before each call, it's cheap) and put it in your agent's context. It returns:

recommended_script: a confirmed test winner, otherwise the best-performing script with enough calls, and why.
do_more / do_less: behaviors that change the meeting rate at 95% confidence, with the rates behind them.
objections: the five most common, how often they turn into meetings, and the answer to use (from the recommended script, or a proven framework).
tests: where each running test stands.

Human and agent calls both feed it, so your agent learns from your best reps, and your reps learn from the agent's volume.