MCP server

One loop, nine tools

Most jobs need one call: get_geo_latest_report returns scores, competitors, recommendations and run deltas, already interpreted.

Quickstart

Five minutes with Claude Code. Other clients are under Connect your client.

1. Check your plan

   MCP comes with Growth and Enterprise. Compare plans

2. Create a Read & write key

   Open Settings, then API keys. Name it after the job, pick Read & write and copy it. It's shown once.

3. Put the key in an environment variable

   Terminal

   export CONTENTSTACK_CANOE_API_KEY="sk_your_key_here"

4. Add the server

   Terminal

   claude mcp add --transport http \
     contentstack-canoe \
     https://canoe.contentstack.com/api/mcp \
     --scope user \
     --header \
     "Authorization: Bearer $CONTENTSTACK_CANOE_API_KEY"

   claude mcp list should show contentstack-canoe as connected.

5. Ask your first question

   Prompt

   List my Contentstack Canoe brands and their monitors. Which monitor lost the most visibility since its last run?

   Reading data never spends prompts.

API keys

Every request sends a key in the header. Keys work on Growth and Enterprise.

Format

sk_..., sent as Authorization: Bearer sk_... or x-api-key: sk_...

Identity

Acts as its creator, in one workspace. If they lose access, so does the key.

Limit

Up to 3 active keys per person in each workspace.

Lifetime

No expiry. Revoke one and it stops at once.

Requests

10,000 a month on Growth, shared with the API. Over MCP, only tool calls count, including failed ones.

Prompts

Only trigger_geo_test_run spends prompts.

Connect your client

Every config uses contentstack-canoe as the server key and keeps the key out of files you commit.

{
  "mcpServers": {
    "contentstack-canoe": {
      "type": "http",
      "url": "https://canoe.contentstack.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${CONTENTSTACK_CANOE_API_KEY}"
      }
    }
  }
}

{
  "mcpServers": {
    "contentstack-canoe": {
      "url": "https://canoe.contentstack.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${env:CONTENTSTACK_CANOE_API_KEY}"
      }
    }
  }
}

{
  "inputs": [
    {
      "type": "promptString",
      "id": "contentstack-canoe-key",
      "description": "Contentstack Canoe API key",
      "password": true
    }
  ],
  "servers": {
    "contentstack-canoe": {
      "type": "http",
      "url": "https://canoe.contentstack.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${input:contentstack-canoe-key}"
      }
    }
  }
}

{
  "mcpServers": {
    "contentstack-canoe": {
      "serverUrl": "https://canoe.contentstack.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${env:CONTENTSTACK_CANOE_API_KEY}"
      }
    }
  }
}

[mcp_servers.contentstack-canoe]
url = "https://canoe.contentstack.com/api/mcp"
bearer_token_env_var = "CONTENTSTACK_CANOE_API_KEY"

{
  "mcpServers": {
    "contentstack-canoe": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://canoe.contentstack.com/api/mcp",
        "--header", "Authorization:${CONTENTSTACK_CANOE_AUTH}"
      ],
      "env": {
        "CONTENTSTACK_CANOE_AUTH": "Bearer sk_your_key_here"
      }
    }
  }
}

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: mcp-client-2025-11-20" \
  -H "content-type: application/json" \
  -d @- <<EOF
{
  "model": "claude-opus-5",
  "max_tokens": 16000,
  "mcp_servers": [{
    "type": "url",
    "url": "https://canoe.contentstack.com/api/mcp",
    "name": "contentstack-canoe",
    "authorization_token": "$CONTENTSTACK_CANOE_API_KEY"
  }],
  "tools": [{
    "type": "mcp_toolset",
    "mcp_server_name": "contentstack-canoe",
    "configs": { "trigger_geo_test_run": { "enabled": false } }
  }],
  "messages": [{ "role": "user", "content": "Summarize the latest Contentstack Canoe report for Tarnwell Cycles." }]
}
EOF

Any client that speaks Streamable HTTP and sends a custom header should work. To check the connection by hand, run npx @modelcontextprotocol/inspector, pick Streamable HTTP and add the header.

Tool reference

All nine, in the order agents usually call them. Every tool but trigger_geo_test_run is annotated read only, so clients that honor annotations run them without asking.

List my Contentstack Canoe brands.

Show every Tarnwell Cycles monitor with its last score, its trend and when it last ran.

Summarize the latest results for our commuter e-bike monitor. Are the numbers from a finished run?

How have our score and brand rank moved over the last 8 runs, engine by engine?

Show me the exact answers where Aldren was recommended and Tarnwell Cycles wasn't mentioned.

Compare how ChatGPT answered "best commuter e-bike under $2,000" in the first September run and in the latest one.

Which third-party sites do the engines cite in the Consideration stage, and which of them also name Solvik or Pellow?

Did our own domain's share of citations grow between July and September?

We just published the Aldren comparison page. Re-run the monitor and tell me what changed.

Reading responses

Fields are snake_case and can be added over time. Check these four before you repeat a number.

run_status

idle, running, success or error. Mid-run, you see the last completed run. has_partial_failure and errored_sections flag what didn't finish.

_guidance.notes

Notes from the tool, not an engine: stale numbers, no runs yet, missing sentiment. Pass them on.

run_deltas.comparable

False with no previous run, so a zero delta means no data. leaderboard_rank_delta is positive when you climbed.

History tiers

Only the latest run has full competitor detail. Older runs keep leaderboard rank.

{
  "run_status": { "status": "success", "last_run_at": "2026-10-05T09:14:02Z", "has_partial_failure": false },
  "identity": { "brand": "Tarnwell Cycles", "product": "Commuter e-bikes" },
  "overview": {
    "overall_score": 67,
    "visibility_pct": 62,
    "brand_rank": { "rank": 3, "of": 10 },
    "sentiment": { "label": "Positive", "score": 74 }
  },
  "funnel": { "awareness": 47, "consideration": 57, "decision": 81 },
  "site_audit": { "readiness_score": 80, "priority_fixes": ["..."] },
  "recommendations": [
    { "rank": 1, "category": "competitive", "title": "Close the gap with Aldren", "how": ["..."] }
  ],
  "run_deltas": { "comparable": true, "score_delta": 6, "leaderboard_rank_delta": 1 }
}

Scores, mentions and citations are computed in code, so report the values as returned. Sentiment is a model's read. Metrics defines every number.

Add the Contentstack MCP

With both servers, your agent reads a recommendation here, finds the entry in your stack and drafts the change for your approval.

{
  "mcpServers": {
    "contentstack-canoe": {
      "url": "https://canoe.contentstack.com/api/mcp",
      "headers": { "Authorization": "Bearer ${env:CONTENTSTACK_CANOE_API_KEY}" }
    },
    "contentstack": {
      "command": "npx",
      "args": ["-y", "@contentstack/mcp"],
      "env": { "CONTENTSTACK_API_KEY": "<YOUR_STACK_API_KEY>", "GROUPS": "cma" }
    }
  }
}

The Contentstack MCP has its own auth: see the Contentstack MCP server docs (opens in a new tab). Then add the Fix it in Contentstack skill, or run it unattended with Contentstack Agent OS.

Security and least privilege

A key carries its creator's authority in one workspace, and the tool list isn't a security boundary. Keep mistakes small.

• One key per integration, named for its job, so you can revoke one alone.
• Keys live in env vars or a secret manager, never in files or chats.
• Block trigger_geo_test_run for read-only jobs. Claude.ai connectors and the Claude API can turn it off. Elsewhere, keep approval on.
• A person approves CMS changes. Contentstack Canoe never writes to your site.
• Revoke quiet keys. Check Last used.

We log each call's tool, user and workspace, never keys, headers or payloads. An id from another workspace returns nothing.

Errors

Tool errors

A failed tool returns isError: true, with code and message as JSON in its first text block.

Out of prompts? No tool error says so: the run's run_status.status turns to error. Check Usage in Settings and Plans and limits.

HTTP errors

These arrive before any tool runs, so clients usually show a failed connection.

Coming soon

• OAuth sign-inComing soonConnect ChatGPT and Claude.ai without a key header.
• Read only keys for MCPComing soonRead-only agent access, enforced by the key.
• Native alertsComing soonScore and competitor alerts without polling. Today, use the Visibility alerts skill.
• Outbound webhooksComing soonEvents like "run finished" sent to your endpoint.