Botlington Agent Token Audit#

Botlington audits AI agents for token waste. Gary (Botlington's AI) runs a 7-question consultation, scores your agent across 5 dimensions, and returns a prioritised list of fixes with estimated monthly savings.

Live endpoint: https://botlington.com/a2a

Agent Card: https://botlington.com/.well-known/agent.json

Pricing: €149/audit — buy at https://botlington.com/checkout

Sample audit: https://botlington.com/audits/stripe

---

Getting an API Key#

1. Go to https://botlington.com/checkout
2. Complete payment (€149 single / €349 for 3 / €749 for 10)
3. Success page returns your api_key

Set it in your environment or pass as x-api-key header.

---

Protocol: JSON-RPC 2.0 over HTTPS#

All requests are POST https://botlington.com/a2a with:

• Content-Type: application/json
• x-api-key: YOUR_API_KEY

Method: tasks/send#

Start a new audit (no taskId = new session):

json

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tasks/send",
  "params": {
    "message": {
      "role": "user",
      "parts": [{ "kind": "text", "text": "start" }]
    }
  }
}

Gary responds with question 1 and a taskId:

json

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "id": "TASK_ID",
    "status": { "state": "input-required" },
    "artifacts": [{
      "name": "gary-question",
      "parts": [{ "kind": "text", "text": "Hi. I'm Gary Botlington IV — I audit AI agents' token usage. ..." }]
    }]
  }
}

Continue conversation (include taskId):

json

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tasks/send",
  "params": {
    "id": "TASK_ID",
    "message": {
      "role": "user",
      "parts": [{ "kind": "text", "text": "I run 8 cron jobs, firing every 15–60 minutes." }]
    }
  }
}

Repeat for each of Gary's 7 questions. On the final answer, state transitions to completed.

Method: tasks/get#

Poll for status after submitting the final answer:

json

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tasks/get",
  "params": { "id": "TASK_ID" }
}

---

Direct Config Submission (Legacy)#

Skip the conversation — submit your config directly:

json

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tasks/send",
  "params": {
    "message": {
      "role": "user",
      "parts": [{
        "kind": "data",
        "data": {
          "agentConfig": {
            "cronJobs": [
              {
                "name": "inbox-scan",
                "schedule": "*/30 * * * *",
                "model": "claude-sonnet-4",
                "systemPrompt": "Check email for urgent messages. Load full inbox context.",
                "tools": ["gmail", "browser", "notion"]
              }
            ],
            "primaryModel": "claude-sonnet-4",
            "contextStrategy": "full-file-reads",
            "toolSurface": ["gmail", "browser", "notion", "slack"]
          }
        }
      }]
    }
  }
}

---

Audit Result Format#

json

{
  "score": 62,
  "grade": "C",
  "summary": "Significant token waste identified across model selection and context strategy.",
  "findings": [
    {
      "id": "finding-001",
      "severity": "critical",
      "dimension": "model-efficiency",
      "description": "3 cron jobs using claude-sonnet for pattern-matching tasks haiku handles fine.",
      "recommendation": "Downgrade mechanical crons to haiku. Reserve sonnet for judgment tasks.",
      "estimatedSaving": {
        "tokensPerRun": 8400,
        "percentReduction": 73
      }
    }
  ],
  "estimatedMonthlySavings": {
    "tokensReduced": 2100000,
    "percentReduction": 41,
    "euroEstimate": 42
  },
  "priorityActions": [
    "Downgrade 3 mechanical crons from sonnet → haiku",
    "Replace full-file context reads with targeted memory queries",
    "Replace browser-based Slack reads with direct API calls"
  ]
}

---

SSE Streaming (GET)#

Stream results as they arrive:

bash

curl -N "https://botlington.com/a2a?taskId=TASK_ID"

Events:

• event: finding — individual finding as it's scored
• event: complete — full result object
• event: working — still processing

---

The 5 Scoring Dimensions#

1. Model efficiency — right model for the task? (haiku vs sonnet vs opus)
2. Context hygiene — loading only what's needed per run?
3. Tool surface — any browser calls replaceable with direct APIs?
4. Prompt density — clear, tight prompts or verbose/ambiguous ones?
5. Idempotency — tracking what's already been done to avoid repeat work?

---

Complete Shell Example#

bash

API_KEY="your-api-key"
BASE="https://botlington.com/a2a"

# 1. Start audit
RESPONSE=$(curl -s -X POST $BASE \
  -H "Content-Type: application/json" \
  -H "x-api-key: $API_KEY" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tasks/send","params":{"message":{"role":"user","parts":[{"kind":"text","text":"start"}]}}}')

TASK_ID=$(echo $RESPONSE | python3 -c "import sys,json; print(json.load(sys.stdin)['result']['id'])")
QUESTION=$(echo $RESPONSE | python3 -c "import sys,json; print(json.load(sys.stdin)['result']['artifacts'][0]['parts'][0]['text'])")

echo "Task: $TASK_ID"
echo "Gary: $QUESTION"

# 2. Answer Gary's question
curl -s -X POST $BASE \
  -H "Content-Type: application/json" \
  -H "x-api-key: $API_KEY" \
  -d "{\"jsonrpc\":\"2.0\",\"id\":2,\"method\":\"tasks/send\",\"params\":{\"id\":\"$TASK_ID\",\"message\":{\"role\":\"user\",\"parts\":[{\"kind\":\"text\",\"text\":\"I run 8 cron jobs, every 15-60 minutes.\"}]}}}"

# ... continue for all 7 turns ...

# 3. Stream results
curl -N "$BASE?taskId=$TASK_ID"

---

Notes#

• One audit credit = one completed 7-turn consultation
• Credits are deducted at conversation start (turn 0), not on completion
• A resumed conversation (same taskId) does not consume additional credits
• If Gary is mid-conversation and you restart with the same taskId, it continues from where it left off
• The agent card at /.well-known/agent.json enables A2A-compatible orchestrators to auto-discover Botlington