Usage

Track API usage, credit consumption, and per-agent AI spend.

Overview

Two usage streams roll up onto the dashboard’s Usage tab and into the /v1/usage API:

• Per-call usage — every successful and failed call to /v1/data/*, /v1/connections/*, /v1/db/* writes a row to usage_logs with the provider id, the endpoint, latency, status, and cost in data credits.
• Per-session AI usage — every deployed-agent session writes one or more agent_compute_usage rows with token counts, cache hits, runtime ms, and cost in AI credits.

Dashboard

Open Usage in the dashboard sidebar. The page shows:

• Daily credit consumption for the data and AI wallets, side by side.
• Per-agent breakdown — which deployed agents are burning the most AI credits, with a click-through to the agent’s session list.
• Top providers — which data APIs and connections are taking the most data credits.
• Recent calls — the live tail of usage_logs with status, latency, and the exact MCP tool name if the call came from an agent.

You can also drill into any single deployed agent from Agents → <agent> to see only that agent’s usage, both data and AI.

Usage API

Pull the same data programmatically:

curl -X GET "https://api.agntdata.dev/v1/usage" \
  -H "Authorization: Bearer YOUR_API_KEY"

Parameters

Parameter	Type	Description
limit	integer	Max rows to return (default 50).
offset	integer	Pagination offset.
providerId	string	Filter to one data source or connection (e.g. linkedin, stripe).
status	string	success or error.
from	string	ISO 8601 start.
to	string	ISO 8601 end.

Example

curl -X GET "https://api.agntdata.dev/v1/usage?from=2026-04-01&to=2026-04-30&limit=100" \
  -H "Authorization: Bearer YOUR_API_KEY"

Response:

{
  "success": true,
  "data": [
    {
      "id": "…",
      "providerId": "linkedin",
      "endpoint": "/get-company-details",
      "costCents": 1,
      "status": "success",
      "createdAt": "2026-04-15T10:30:00Z"
    }
  ]
}

When a row originates from a deployed agent, the metadata object includes tool_name (the exact MCP tool name the agent called) and deployed_agent_id so you can correlate API spend back to a specific agent.

Agent compute usage

For AI credits specifically, each session row carries the breakdown:

Field	Meaning
input_tokens	Tokens sent to the model in this turn or aggregated across the session.
output_tokens	Tokens the model generated.
cache_read_tokens / cache_write_tokens	Prompt-caching activity.
runtime_ms	Wall-clock time the managed runtime took.
cost_cents	Total deducted from the AI wallet for this row.

You can inspect this from inside the Builder via get_agent_session_usage (per session) or from the dashboard’s agent detail page (per agent, rolled up over time).

Best practices

Cache aggressively

Repeat lookups against the same identifier within a short window are good candidates for application-level caching. The data APIs include latency hints but they don’t dedupe across requests.

Right-size endpoints

Light-touch endpoints (e.g. profile basics) cost less than deep enrichment. Pick the smallest endpoint that gets you the data you actually use.

Prune unused agent tools

Every tool you expose to an agent is loaded into its context window and counts against input tokens on every turn. Strip tools the agent never uses to lower AI credit burn.

Monitor regularly

The Usage tab is the cheapest way to spot regressions. A sudden spike in one provider’s calls usually points at a logic bug, not a feature win.

Next steps

Credits

Data wallet vs. AI wallet, auto-reload, balance checks.

Rate limits

Per-plan RPM and backoff.

API keys

Manage workspace API keys.