Live // Elastic Observability

AI Operations

My own AI usage, as it happens.
Every model request, token and tool call from my Codex and Claude Code sessions, and every API call I send through OpenRouter, sent to Elastic as I work. It’s the observability I set up for clients, pointed at myself. Prompts and responses are never captured.

Loading…

Selected window // Last 30 days

Codex · model requests

—

Codex · output tokens

—

Claude Code · model requests

—

Claude Code · output tokens

—

OpenRouter · model requests

—

OpenRouter · output tokens

—

Bars: each tool’s share of the three together.

Output tokens over time

Tokens the models wrote for each tool, a day to a bar: the code, commands and answers they produced, not the context they read.

CodexClaude CodeOpenRouter

Latest telemetry Last event

The latest event received from each tool
Codex—
Claude Code—
OpenRouter—

Model request time Avg · slowest

How long model requests took, by tool
ToolRequestsAverageSlowest
Codex———
Claude Code———
OpenRouter———

Tokens by model

The five models with the most tokens in the window. Input counts the cached context each request re-reads, so it runs far ahead of output. Each column has its own scale. Token totals measure usage, not what any of it cost.

Codex Input incl. cache · output

    Claude Code Input incl. cache · output

      OpenRouter Input incl. cache · output

        What we mask

        Before Elastic stores a record, a pipeline there removes or hashes the details that could name a client or leak a secret: account IDs, folder and repository names, keys and tokens, email addresses in text, URLs past the domain. The same folder always gets the same code, so the data still groups. Counted on traces and events since October 19, 2026.

        Masked before storage Records

        How many records had each kind of detail masked
        What was maskedRecords

        In this window Count

        Masking totals in the selected window
        Details masked—
        Records with something masked—
        Records checked—

        One request, as Elastic stored it

        A recent Claude Code request of mine that used a tool, span by span. Open a span to see its fields. ******** stands for a value that was removed, hashed, redacted or scrubbed before storage, or never captured at all. Other fields stay in Elastic.

          How it works

          The same pieces I put in place for a client’s team, at the size of one person.

          1. Sent as I work

            Codex and Claude Code send OpenTelemetry traces and metrics straight to Elastic Cloud, and OpenRouter sends a trace of each API call. No collector in between.

          2. Content stays off

            Prompt text, responses, tool arguments and file contents are switched off at the source. Elastic gets counts, models, tool names and timings, and masks identifying details before storing them.

          3. Totals and one sample

            This page asks Elastic for totals and one recent request with a read-only key, at most once a minute, and refreshes while it’s open. The sample shows a fixed list of fields, the names of masked ones, and no IDs.

          Want this view of your team’s AI? It’s the “Your dashboards” clause in every scope: results tracked in your own account.

          Book a 20-minute call