Developers · MCP

Drive Summit from your coding agent

The Summit MCP server plugs Summit into any MCP-aware coding agent — Claude Code, Codex, Gemini CLI, Cursor. Hand it a page and it returns a selector-precise fix checklist; on Pro, it drives your whole workspace: approve findings, launch A/B tests, and read the Bayesian verdict — all without leaving your editor.

What the agent can do

Two tiers. The audit tools need nothing but a URL — they work off a public share token, no account. The workspace tools drive your Summit workspace and need a Pro API token.

Audit tools Free · no login

Tool What it does
summit_run_audit(url, email) Kick off a new audit (1 free per email). Returns a share link to poll.
summit_get_audit(report) Score, grade, screenshot, and ranked findings for a share token or link.
summit_implementation_plan(report) Turn the audit into an ordered, code-ready checklist — selector, before→after, lift.
summit_list_findings(report, tier) Findings filtered to one tier: must_fix / should_fix / nice_to_fix.

Workspace tools Pro

The full Study → Approve → Ship loop. Reading is safe; the mutating tools change workspace state but never touch your live site directly — variants only serve after an experiment is explicitly launched.

Tool What it does
summit_workspace_overview() KPI rollup: 7-day visitors/conversions, running experiments, pending reviews, winners shipped.
summit_review_queue() Everything waiting on human sign-off — proposed findings + built experiments.
summit_list_sites() Sites in your workspace.
summit_list_experiments(site_id) Experiments with status + winner.
summit_experiment_results(experiment_id) Bayesian verdict: leader, lift, P(beat control), significance.
summit_site_pulse(site_id) Snippet install check + 7-day visitors / conversions / rage clicks.
summit_approve_finding(finding_id) · mutates Approve a fix → builds an A/B experiment + variants.
summit_reject_finding(finding_id) · mutates Dismiss a proposed fix.
summit_approve_experiment(experiment_id) · mutates Approve a built experiment for launch.
summit_launch_experiment(experiment_id) · mutates Start serving the A/B test to live traffic.

Install the server

The MCP server is a small stdio program that talks to the Summit API. Install it once; any MCP-aware client can then launch the summit-mcp command.

shell 
# From the Summit repo's mcp-server/ directory:
uv sync                    # uses the pinned uv.lock
# …or, without uv:
pip install -e . -c requirements.txt

# installs the `summit-mcp` command (= python -m summit_mcp)

Connect your workspace

The workspace tools authenticate with a workspace API token (SUMMIT_API_TOKEN) scoped to a single workspace (SUMMIT_WORKSPACE_ID). Generating a token is a Pro feature — here’s your status:

Checking your plan…

Register with your agent

Point any MCP client at the summit-mcp command over stdio. Drop in your token and workspace ID from the panel above.

Codex CLI

toml 
# ~/.codex/config.toml
[mcp_servers.summit]
command = "summit-mcp"
env = { SUMMIT_API_BASE_URL = "https://api.trysummit.ai", SUMMIT_API_TOKEN = "smt_your_token_here", SUMMIT_WORKSPACE_ID = "your_workspace_id" }

Gemini CLI

json 
// ~/.gemini/settings.json
{
  "mcpServers": {
    "summit": {
      "command": "summit-mcp",
      "env": {
        "SUMMIT_API_BASE_URL": "https://api.trysummit.ai",
        "SUMMIT_API_TOKEN": "smt_your_token_here",
        "SUMMIT_WORKSPACE_ID": "your_workspace_id"
      }
    }
  }
}

The Claude Code one-liner is pre-filled with your workspace ID in the Connect panel above.

Example flows

Audit → fix (free)

You: Audit https://moonsign.co.in and fix the top 3 conversion issues.

Agent → summit_run_audit(url="https://moonsign.co.in", email="me@co.com")
      ← { share_url: ".../audit?r=TOK", status: "queued" }
Agent → summit_implementation_plan(report="TOK")     # poll until completed
      ← { steps: [ { selector: "a.cta", before: "Learn more",
                     after: "Get my free reading", expected_lift_pct: 12 }, … ] }
Agent then edits the codebase per each step.

Drive the whole loop (Pro)

You: Anything waiting on me? Approve the highest-lift fix and launch it.

Agent → summit_review_queue()
      ← { findings: [ { id: "…", title: "Weak hero CTA", estimated_lift: 14 } ] }
Agent → summit_approve_finding(finding_id="…")        # builds experiment + variants
Agent → summit_approve_experiment(experiment_id="…")  # after QA
Agent → summit_launch_experiment(experiment_id="…")
      … later …
Agent → summit_experiment_results(experiment_id="…")
      ← { winner_key: "b", leader_prob: 0.97, is_significant: true }

Keep tokens safe