CodeBurn: Best Developer Analytics CLI Tools for AI devs in 2026

What Is CodeBurn?

CodeBurn is one of the best Developer Analytics CLI tools for AI devs. Built by getagentseal, CodeBurn is a Node.js 20+ terminal app that reads local session data from Claude Code, Codex, Cursor, OpenCode, Pi, and GitHub Copilot and turns it into token, cost, and retry analytics. It is for people who want to see where AI coding spend goes, not another SaaS collector or billing proxy. The repo supports six providers, exports JSON and CSV, and tracks one-shot success rate by activity type.

Quick Overview

CodeBurn is local-first, so the report is derived from files already on disk instead of a remote usage API. That makes it useful for private workstations, air-gapped laptops, and teams that do not want assistant telemetry leaving the machine.

Who Should Use CodeBurn?

• Solo indie hackers shipping with Claude Code or Cursor who need a fast read on daily token burn across projects.
• Platform and engineering leads who want to compare usage by model, tool, MCP server, and project without chasing manual self-reports.
• Cost-conscious teams trying to spot edit, test, and fix loops, because CodeBurn highlights one-shot success rate and the most expensive sessions.
• Power users who want jq-friendly JSON for scripts, dashboards, or internal audits.

Not ideal for:

• Teams that do not store assistant sessions locally on the developer machine.
• Organizations that need centralized server-side telemetry across a fleet.
• People who only want a cloud billing portal and do not care about terminal workflows.

Key Features of CodeBurn

• Local session ingestion — CodeBurn reads session data directly from ~/.claude/projects/, ~/.codex/sessions/, ~/.pi/agent/sessions/, Cursor's state.vscdb, OpenCode SQLite, and GitHub Copilot state. There is no wrapper, no proxy, and no API key management.
• Provider plugin system — Each provider adapter handles discovery, JSONL parsing, tool normalization, and display names. The repo says adding a new provider is a single-file job, which keeps the core CLI small and easy to extend.
• Terminal dashboard — The default interface is an interactive TUI with gradient charts, responsive panels, arrow-key navigation, and shortcuts like 1 through 5 plus c for model comparison. It gives you a terminal-native view instead of another browser tab.
• Cost and token breakdowns — CodeBurn groups usage by task type, tool, model, MCP server, project, and shell command. It also shows average cost per session and the five most expensive sessions, which helps isolate retry-heavy work.
• One-shot success rate tracking — CodeBurn measures whether a task type succeeds on the first attempt or burns tokens on edit, test, and fix retries. That turns raw spend into a signal about workflow quality.
• JSON and CSV export — report, today, month, status, and export can emit structured output for jq, spreadsheets, or CI jobs. The JSON payload includes overview metrics, daily breakdowns, projects, models, activities, core tools, MCP servers, and shell commands.
• Filtering and time windows — The CLI supports --provider, --project, --exclude, --from, and --to so you can slice usage by assistant, repository, or exact date range. That makes CodeBurn usable for both daily checks and monthly audits.

CodeBurn vs Alternatives

Pick OpenTrace when you need to see step-by-step execution data for AI workflows. CodeBurn is the better fit when the question is where tokens and money went on the local machine.

Pick Claude Context Mode when your main problem is prompt-window discipline inside Claude. CodeBurn is for post-hoc cost analysis, while Claude Context Mode is for shaping the working context during the session.

Pick OpenSwarm when you are coordinating multiple agents or tasks in parallel. CodeBurn does not orchestrate agents; it tells you which assistants, projects, and tool calls are burning budget.

How CodeBurn Works

CodeBurn uses provider adapters to read session files directly from disk, normalize events into a shared model, and merge them into one report. That model covers task type, tool name, model, MCP server, project, shell command, and session metadata, so the dashboard can compare Claude Code, Codex, Cursor, OpenCode, Pi, and GitHub Copilot without special cases in the UI.

The pricing layer comes from LiteLLM, which the project caches so all supported models can be costed without calling a remote billing endpoint. Cursor and OpenCode need extra handling because they store data in SQLite, while Claude Code, Codex, Pi, and Copilot rely on file-based session formats. The result is a local scan that can run offline once the data exists.

codeburn report --provider claude --project myapp --format json | jq '{cost: .overview.cost, topProjects: .projects[0:5]}'

This command limits the scan to Claude Code sessions for myapp, emits structured JSON, and trims the result down with jq. Expect the first scan of a large Cursor or OpenCode database to take longer, because CodeBurn caches derived data after the initial pass.

Pros and Cons of CodeBurn

Pros:

• Reads assistant session data from the local filesystem and local SQLite databases, so it does not need a proxy or browser extension.
• Normalizes multiple providers into one cost model, which makes cross-tool comparisons practical.
• Tracks one-shot success rate, so you can see which workflows waste tokens on retries.
• Exports JSON and CSV, which makes it easy to feed CodeBurn into scripts, spreadsheets, or internal tooling.
• Supports project and date filtering, which is useful for audits, retrospectives, and budget checks.
• Has a terminal UI with keyboard shortcuts, so repeated checks are quick for shell-oriented developers.

Cons:

• Only sees sessions that exist on disk, so it cannot analyze work that never lands in local provider storage.
• Cursor Auto mode uses estimated Sonnet pricing, which means some costs are approximate rather than exact.
• GitHub Copilot only exposes output tokens in the session state, so cost rows can sit below actual API spend.
• The first scan of large SQLite-backed providers can be slow before cached results are available.
• It is an individual-machine analytics tool, not a centralized org-wide telemetry platform.

Getting Started with CodeBurn

npm install -g codeburn
codeburn
codeburn report --format json | jq '.overview'

That installs CodeBurn globally and opens the default dashboard, which starts on a 7-day view. If you prefer not to install anything, npx codeburn works too, and you can narrow the view later with codeburn today, codeburn month, or codeburn report --provider cursor.

The first run needs Node.js 20+ and whatever provider session directories already exist on your machine. If you use Cursor or OpenCode, CodeBurn may install better-sqlite3 automatically as an optional dependency, then cache the derived scan so later runs are much faster.

Verdict

CodeBurn is the strongest option for local AI usage audits when your assistants already write session data to disk and you want cost plus retry analytics in the terminal. Its main strength is provider-wide normalization; its caveat is that it cannot see sessions that never land on disk. Use CodeBurn if you want a machine-local answer to where your AI tokens went.