Coding agents

Claude Code, Codex and Antigravity

Coding agents already record every model call they make. AgentCost reads that record, so their sessions show up next to the rest of your LLM spend with no instrumentation, and your guardrails can stop a call before it runs instead of reporting it afterwards.

Connect a machine

One command per agent, once per machine. It imports the last 30 days and installs a hook, so each finished turn is sent as it happens, from any folder, in the terminal, the IDE extension or the desktop app.

bash

pip install agentcost

agentcost sync --dry-run                       # look first: no key, nothing sent
agentcost connect claude-code --api-key sk_your_project_api_key
agentcost connect codex
agentcost connect antigravity

Codex asks you to trust a new hook: open Codex and run /hooks once. Antigravity has no hooks, so its usage is sent whenever another connected agent syncs, or when you run agentcost sync on a schedule.

What you get

A session is a run. Open one to see every call in it, or group spend by model, tool, subagent, session or developer.

	Claude Code	Codex	Antigravity
Cost per session, model and developer	Yes	Yes	Estimated
Cost per tool	Yes	Yes	No
Subagents as their own agents	Yes	When the log names them	No
Syncs by itself after every turn	Yes	Yes, once the hook is trusted	With another agent, or on a schedule
OpenTelemetry, nothing installed	Yes	Yes	No
Tool and model boundaries enforced	Yes	Yes	No
Per-run limits and budget enforced	Yes	Yes	No

Antigravity publishes no usage export. Its figures are read from an undocumented local format and priced from the model name it records, so treat them as estimates. Conversations from before June 2026 are encrypted and cannot be read.

Subscriptions are not spend

On a Claude or ChatGPT plan, or in Antigravity, nothing is billed per token. Those calls are still priced, so you can see what the usage is worth at API rates, but each one is marked billing: subscription and is left out of your budget, its alerts and its hard cap. The overview shows the two apart. Usage on an API key is metered and counts as spend.

Enforce guardrails

Add --enforce and a second hook, the gate, runs before every tool call and every prompt. It refuses what the project does not allow and tells the agent why.

bash

agentcost connect claude-code --enforce

It enforces the guardrail you declare for the agent named claude-code or codex on the Guardrails page. A subagent answers to claude-code/<type> if you declare one, and to the main agent otherwise.

Allowed tools: A tool that is not on the list is refused.
Read-only: A tool tagged as writing is refused.
Allowed models: Tool calls and prompts are refused while the session is on a model that is not on the list.
Max tool calls per run: Further tool calls in that session are refused.
Max cost per run: Further tool calls and prompts in that session are refused.
Monthly budget, hard cap: New prompts are refused once it is spent. Metered usage only, never a subscription.

The gate never calls the network. It decides from a policy the sync keeps current and from the session's own log, adds about a fifth of a second to a tool call, and applies a changed guardrail from the next turn. When it cannot decide, it allows the call.

A whole organisation, with nothing installed

Claude Code and Codex can export their own telemetry. Point it at AgentCost from managed settings and every session reports in, including ones that run in the cloud and leave no local log. Only the per-call usage records are read; prompts, tool results and everything else in the export are discarded unread.

Claude Code: settings.json

{
  "env": {
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_LOGS_EXPORTER": "otlp",
    "OTEL_METRICS_EXPORTER": "none",
    "OTEL_EXPORTER_OTLP_PROTOCOL": "http/json",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "https://api.agentcost.tech/v1/otlp",
    "OTEL_EXPORTER_OTLP_HEADERS": "Authorization=Bearer sk_your_project_api_key",
    "OTEL_RESOURCE_ATTRIBUTES": "agentcost.billing=subscription"
  }
}

Codex: config.toml

[otel]
exporter = { otlp-http = {
  endpoint = "https://api.agentcost.tech/v1/otlp/v1/logs",
  protocol = "json",
  headers = { "Authorization" = "Bearer sk_your_project_api_key" }
}}

State how the usage is paid for with agentcost.billing=subscription or api; Claude Code does not say, and unstated usage counts as metered. A machine that reports both ways is counted once. This path tracks only: enforcement needs the hook.

Let the agent read its own spend

Add the MCP server with your project key and the agent can check the budget, or what a run cost, before it starts something expensive.

bash

claude mcp add --transport http agentcost https://agentcost.tech/api/mcp \
  --header "Authorization: Bearer sk_your_project_api_key"

What leaves the machine

Token counts, model, tool names, session id and timestamps. Prompts, replies, tool arguments and tool output are never sent or kept. The working folder's name and git branch are sent only if you pass --include-repo, and never the path or the remote URL.

Related: CLI reference · MCP server · Data & privacy