# Trace Claude Code, Codex and other coding-agent sessions and their cost

> Trace Claude Code, Codex and other coding-agent sessions and their cost: 5 companies fully do it, 10 with a limit, 22 carry the tag. Every listing in Agent observability & evals with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-observability.coding-agent-sessions/
- Markdown: https://www.anchorterminal.com/niches/agent-observability.coding-agent-sessions/index.md (~2,050 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-observability.coding-agent-sessions/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-observability.coding-agent-sessions/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 5 | 10 | 22 | 7 | 8 |

## What counts as done

Every session appears as a trace with its turns, tool calls and sub-agents, token usage per session, and a session id the team can cite, with no proxy in the agent's network path and no silent loss of sessions.

How agents ask for it: what did Claude Code do in that session; which sub-agent spent the tokens; trace my Codex runs for the team; connect Claude Code on a subscription plan to a trace store

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Arize Phoenix (Arize AI, Inc.)**: The coding-harness-tracing plugin hooks Claude Code lifecycle events and Codex notify events: one trace per turn with LLM, nested tool and subagent spans, token cost, grouped by session, no proxy. The Agent SDK needs ClaudeSDKClient, not query().
- **Braintrust API + MCP (Braintrust Data, Inc.)**: bt trace enable claude or codex installs plugins that record session, turn, model, tool and subagent spans with no proxy; Claude Code's native OTel export is the alternative. The coding agent insights dashboard breaks cost down by user, key, model and branch.
- **Laminar API + MCP (LMNR AI, Inc.)**: lmnr-cli installs Claude Code and Codex plugins through the harnesses' own marketplaces; each turn is a trace with LLM calls, tool executions and subagents grouped as a session, with a project key in a config file and no proxy. The Agent SDK is wrapped in code.
- **LangSmith API + MCP (LangChain, Inc.)**: Plugins for Claude Code and Codex send each turn as a trace grouped into a thread, with managed settings for fleets; no proxy. Tracing stops and reminds the user in chat when a sign-in expires, so gaps are not silent.
- **Respan API + MCP (Keywords AI Inc. (d/b/a Respan))**: The session-traces plugin runs inside Claude Code with no hook script or proxy and sends each turn as a trace with a span per model request and tool call, subagents nested, and token, cache and timing data; a Codex guide covers tool calls and tokens.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agent Tree (No legal entity found. The LICENSE, plugin.json and marketplace.json name "Mike Pollow, Digitalroots")**: Shows a Claude Code session's sub-agents as a live tree with reports, files touched, usage and a log, through hooks and no proxy. Claude Code only, and no trace or session id survives the session.
- **agents-tree (No legal entity found. The LICENSE and pyproject.toml name Karel Blavka)**: Reads Claude Code, Codex, Grok Build and Antigravity CLI transcripts on disk and shows sessions, sub-agents, models, tool calls and tokens with no proxy or hook. Nothing is stored, no cost, and a harness update can blank columns.
- **Helicone AI Gateway + MCP (Helicone, Inc.)**: Claude Code logs through ANTHROPIC_BASE_URL pointed at anthropic.helicone.ai with the key in the URL, a method marked maintained but not developed; Codex uses the gateway base URL. Both put a proxy in the agent's path and record requests, not hooks or subagents.
- **Langfuse API + MCP (Langfuse GmbH (a ClickHouse, Inc. company))**: Claude Code is traced by a Stop-hook script that reaches into SDK 4.x internals and must be pinned; Codex has a plugin with a Stop hook that fails open unless LANGFUSE_CODEX_FAIL_ON_ERROR is set. Sessions, tools and cached tokens are grouped per session_id.
- **MLflow Tracing (MLflow Project, a Series of LF Projects, LLC)**: mlflow autolog claude writes hooks that record tools, nested subagents, skills and per-session token cost; Codex and Gemini CLI use a notify hook, no proxy. The Stop hook once dropped traces silently (issue 25517) and only Claude Code was supported at first (20307).
- **Navige (Navige Ltd)**: Claude Code calls go through ANTHROPIC_BASE_URL with x-navige-key, measured per session and subagent and separating plan usage from paid extra usage, with trimming and caching off. A proxy sits in the path and tool calls are checked, not recorded with content.
- **Opik API + MCP (Comet ML, Inc.)**: The Claude Code plugin captures every turn, tool call and nested subagent with cost, no proxy; the Codex path is Codex's own OTel exporter, which carries operational spans only, with no prompt text, tokens or cost.
- **Overmind (Overmind Technology Ltd (England and Wales, company 16585996))**: Claude Code spans arrive over OTel with OTEL_LOG_USER_PROMPTS=1, but tool input and responses need detailed tracing that only allowlisted organisations get in interactive sessions; no hook or plugin and no Codex path.
- **Pydantic Logfire (Pydantic Services Inc.)**: Codex has two paths: native OTel logs and metrics to /v1/logs, and the Logfire Exporter plugin for turns with tool calls; both need a write token. No Claude Code session integration was found, only the MCP plugin for querying.
- **W&B Weave (Weights and Biases, LLC)**: Agent Lens, named Weave's successor on the same backend, has Claude Code and Codex plugins that trace every turn, tool call and subagent with no proxy; a weave-claude-code plugin must be removed first and no PII scrubbing is implemented.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **LangWatch (Reasoning Engine B.V. (doing business as LangWatch))**: The listing claims Claude Code and Codex session tracking, but the docs index has no integration page for either harness, only Claude Platform OTel and a governance ingestion endpoint for external platforms.

- not this job: BotGauge (BotGauge AI Inc. (Delaware, United States)): No Claude Code or Codex integration in the docs index.
- not this job: DeepEval (Confident AI, Inc.): No Claude Code or Codex integration.
- not this job: Galileo API + MCP (Galileo Technologies, Inc.): No Claude Code or Codex integration in the docs index.
- not this job: Gauge CLI (Gauge): Gauge launches its own coding-agent sessions and keeps their transcripts under .gauge/runs; it does not ingest a team's Claude Code or Codex sessions.
- not this job: HoneyHive (HoneyHive Inc.): The coding agents page equips Cursor or Claude with skills, a CLI and a docs MCP to integrate HoneyHive; it does not trace Claude Code or Codex sessions.
- not this job: Prefactor (Prefactor Pty Ltd): No Claude Code or Codex integration.

## What stops agents finishing

- harness exposes no hooks so a proxy is needed
- subscription plan without an API key
- hook drops traces with no error
- only one harness supported

## Around it

Before: agent-harnesses.pair-in-terminal, agent-harnesses.parallel-worktrees. After: agent-observability.cost-per-run, agent-observability.debug-failed-run.

## Where this comes from

Weight 7 of 100 in Agent observability & evals: 8 of 86 accounts: the LLM proxy thread (OTel from Codex and Claude Code, MLflow session tracing, Laminar's in-process proxy), MLflow plugin issues on silently dropped Stop-hook traces and other CLIs, the Opik Claude Code SDK request, the Phoenix Max-plan thread, the sub-agent cost CLI

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-observability.coding-agent-sessions.json, or the MCP tool get_use_case with id agent-observability.coding-agent-sessions. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
