ccost

Find old Codex and Claude Code sessions in seconds, then see which ones burned the most tokens.

ccost turns local AI coding session logs into a searchable terminal UI with a cost lens. Use it to recover past prompts, inspect expensive runs, and understand which sessions, models, and web-search calls drove your spend.

https://github.com/user-attachments/assets/0869bcde-96be-4d98-9c07-c0586b0ea36a

brew install --cask peterxcli/tap/ccost
ccost

Why use it?

• Find the session where you fixed a bug, explored an API, or hit a useful error.
• Sort sessions by estimated cost, tokens, time, model, web searches, session id, or first prompt.
• Browse Codex and Claude Code logs without uploading prompts, paths, stack traces, or source context.
• Keep startup fast with a persisted search index that only re-indexes changed files.

Supported Sources

Install

brew install --cask peterxcli/tap/ccost

Upgrade:

brew update
brew upgrade --cask ccost

Other install paths are on the roadmap:

• cargo install ccost
• cargo binstall ccost
• Linux release artifacts
• Shell install script

Run

Open the default Codex session directory:

ccost

Choose a source explicitly:

ccost --codex
ccost --claude

Use a custom session directory:

ccost --sessions /path/to/jsonl/sessions

Use it when...

Find an old fix

You remember asking Codex or Claude Code about JWT refresh, but not which project or session it was in.

Press /, type:

JWT refresh

Find expensive sessions

Sort by cost to see which runs used the most estimated spend.

Press:

s

until sort is cost.

Audit web-search-heavy sessions

Sort by web searches to find sessions where search calls contributed to the estimate.

Privacy

ccost is local-first.

• No API key required
• No telemetry
• No network calls
• No uploads
• Session JSONL files remain the source of truth
• Cache files are disposable and stored locally

If cache format drift or corruption is detected, delete the cache folder shown in the TUI. The original session logs are not modified.

Features

• Full-text search across local session logs.
• Prefix matching, match highlighting, and visible search cursor.
• Cost-aware sorting by total cost, time, tokens, web searches, model, session id, or first prompt.
• Codex and Claude Code defaults, plus custom JSONL session directories.
• Built-in pricing table with local JSON override support.
• Optional --no-web-cost mode for token-only estimates.
• Single-writer cache lock with read-only and force-write escape hatches.

Why it is fast

• Rust terminal UI.
• Persisted full-text index.
• FST term index for compact prefix search.
• Persisted Merkle tree and file watcher, so unchanged sessions are reused and changed session files are re-indexed incrementally.

Controls

Options

ccost [--codex | --claude | --sessions PATH] [--pricing PATH] [--no-web-cost] [--read-only-index] [--force-index]

• --codex: open the default Codex session directory.
• --claude: open the default Claude Code project transcript directory.
• --sessions PATH: open a custom Codex or Claude Code session directory containing JSONL files.
• --pricing PATH: load a local pricing JSON override.
• --no-web-cost: disable web-search call cost in estimates.
• --read-only-index: open without writing the persisted search cache.
• --force-index: write without the lock. Use only after confirming no other TUI is running.

Pricing

Cost is an estimate based on the token usage and model data present in local session logs. Built-in pricing includes GPT-5.5, GPT-5.4, Claude Opus/Sonnet/Haiku model families, and web-search defaults. Override the local table with --pricing pricing.json:

{
  "web_search_per_1k": 10.0,
  "models": {
    "gpt-5.5": {
      "input_per_m": 5.0,
      "cache_creation_input_per_m": 0.0,
      "cached_input_per_m": 0.5,
      "output_per_m": 30.0,
      "long_context_threshold": 272000,
      "long_context_input_multiplier": 2.0,
      "long_context_output_multiplier": 1.5
    },
    "gpt-5.4": {
      "input_per_m": 2.5,
      "cache_creation_input_per_m": 0.0,
      "cached_input_per_m": 0.25,
      "output_per_m": 15.0,
      "long_context_threshold": 272000,
      "long_context_input_multiplier": 2.0,
      "long_context_output_multiplier": 1.5
    },
    "claude-sonnet-4-5": {
      "input_per_m": 3.0,
      "cache_creation_input_per_m": 3.75,
      "cached_input_per_m": 0.30,
      "output_per_m": 15.0
    }
  }
}

For older pricing overrides, long_context_multiplier is still accepted and applies to all token classes when the input/output-specific fields are omitted. cache_creation_input_per_m is optional and defaults to 0.0.

Benchmarks

Benchmark numbers should be measured on a named machine and reproducible dataset before being used as proof. The launch checklist in docs/github-launch.md includes the table to fill in once those numbers are available.

How is this different?

Roadmap

• Static README screenshot with sanitized data
• GitHub social preview image
• cargo install ccost
• Linux release artifacts
• Demo fixture or ccost --demo
• JSON/CSV export
• Project-level cost summaries
• Cursor, OpenCode, Gemini CLI, and Goose log support
• Configurable themes
• Benchmarks in CI

Contributing

Contributions are welcome, especially for new log sources and export formats. See CONTRIBUTING.md.

If ccost helps you recover an old session or understand where tokens went, a star helps show it is worth maintaining.