OpenTab

Your AI coding tools keep a tab. OpenTab opens it.

See where your coding agents spent their tokens, time, and money.

OpenTab brings the history your tools already keep into one interactive browser. Follow a busy month into a project, a session, its subagents, and individual turns. See what filled the context window. Compare the same token usage at another model's prices.

Works with Claude Code, Codex, OpenCode, Copilot, and more. Runs locally, in your terminal or web browser. No account, no telemetry, and your agents' history stays read-only.

_{Follow one session. Then zoom out. Walk one session from cost to subagents, tool calls, file changes and context, then every machine and the browser. Synthetic showcase data; click for the full 73-second video.}

Try it on your own history

pipx install opentab-ai
opentab

OpenTab discovers supported tools' local records and brings them together automatically. Start with a recent day, open Sessions, and pick one you remember. Press Enter to drill in, Esc to step back, and ? for the keys that work where you are. The mouse works too.

On a subscription? You can still explore your usage: OpenTab starts with clearly labeled API-equivalent estimates for tokens that have no recorded cost. $ switches to recorded spend. How the money works.

Python 3.9+ · macOS · Linux · WSL · native Windows.

More ways to install and upgrade

The PyPI package is opentab-ai; the command is opentab.

brew install hamidi-dev/tap/opentab         # Homebrew (macOS / Linux)
pip install --user opentab-ai               # pip
curl -fsSL https://opentab.hamidi.dev/install.sh | bash

Upgrade with pipx upgrade opentab-ai, brew upgrade opentab, or pip install -U --user opentab-ai.

The runtime uses Python's standard library, with windows-curses added automatically on native Windows. Windows & WSL setup.

New here? Read the Quick start. Prefer a browser? Run opentab web. Showing someone your screen? opentab --demo. A tool missing? opentab doctor. opentab --help is the quick guide. Automating it? See Programmatic access for JSON commands and the local MCP server.

Follow the work behind the numbers

Follow an expensive session down to the turn

Pick a day on the spend heatmap, open its sessions, sort by cost, and follow the largest one. Subagents shows how its cost splits between the main agent and its delegates. Turns lists prompts and the steps after them with tokens and cost; on supported tools you can read a turn's recorded narration, reasoning and exact tool calls. Tools groups usage by tool and MCP server. Sessions · Keys

Understand why token counts and costs tell different stories

A large token count is only the beginning. Token economics puts each token type's share of the volume beside its share of the cost at API list rates: uncached input, output, reasoning, cache reads, and cache writes.

See how much of your usage is cached, and which token types account for the money. Drill into a model within a month or project, or in Trends over the active date range, to see its own economics and the sessions that used it, ranked by that model's contribution.

See what filled the context window

The Context view follows the window as it grows, marks compactions, and estimates what went into it: tool results, call arguments, assistant text, prompts, and more. See when the context shrank and what it grew back to.

_{Two compactions, and the context fills again. The curve shows when; the breakdown shows what was sent. Anonymized demo data.}

Session lists also show Worked time where the records support it: the agent's working bursts, with idle waits for your next prompt removed.

Put another model's prices against your actual usage

Press w, choose a model, and open a session. Compare the list-price cost of the models it used with the cost of those same tokens at your chosen model's rates. Sessions with subagents also show the target price beside each agent's usage.

Explore models you've used or the full bundled catalog. P compares model and provider rates using your token mix, including cache usage, and lets you pin a shortlist. Both sides of a session comparison use list rates; it holds token usage fixed and makes no prediction about another model's output quality or how many tokens it would use.

Detail views depend on what each tool records. Compare support below.

Built to be poked at

Move between daily, weekly and monthly trends, a calendar heatmap, and rankings of models, providers, projects and tools. Follow a chart into the sessions behind it, filter as you type, and change the date range.

• Find your way back. Bookmark sessions, add notes, and reopen supported sessions in their original tool with L, or use opentab launch for an fzf picker. Sessions
• Scope and filter. Filters
• See file edits. The Changes tab shows the edits a session recorded (OpenCode and GitHub Copilot). Changes
• Make it yours. 30 themes, remappable keys, remembered preferences, CSV export. Configuration · Keys

Watch the full narrated tour on YouTube

Search past conversations

Find what you discussed across local OpenCode, Claude Code, Codex, Hermes, pi and omp sessions with Ctrl-F, the CLI or MCP. It is keyword search over recorded user/assistant text. It is off until you enable it: building the index stores conversation text in plaintext on your machine, and you must opt in explicitly. Setup, limits, and clearing the index · MCP configuration

In your terminal. In your browser. In one file.

opentab web                        # open the live browser on localhost
opentab --html report.html          # write a self-contained report
opentab --demo --html demo.html    # anonymize it for sharing
opentab web --html full.html --include-details  # bundle Turns, Tools and Context

The web browser shares the TUI's navigation, themes, trends, and model comparisons. The live version fetches Turns, Tools, and Context on demand where supported. The HTML export packages the overview and comparison views into one file you can open offline, with no server or dependencies. Links to months, days, and sessions let someone open the part you want to show them.

_{Keep exploring in the browser. The same history, with familiar keys and clickable tables.}

Turn traces require --allow-raw-content; static reports also need --include-details or --details-dir. Those detail options include retained user prompts even without traces. Personal notes stay out of web reports. Web browser & exports

Every machine, one tab

Your laptop, workstation, and remote box can all contribute to the same view. opentab pull gathers usage summaries over SSH, with no background agent or listening service and no raw traces in summaries. SSH setup and managing machines

Live prices in your sidebar

Keep an eye on a session while it runs. herdr-opentab puts each running agent's cost beside it in the Herdr sidebar, subagents included.

For your own status bar or script, opentab cost "$PWD" prints the latest session's cost for that project. opentab --goto "$PWD" opens that session for inspection.

See the Herdr integration

_{Amounts synthetic, everything else real.}

The tools it reads

OpenCode · Claude Code · Codex CLI · GitHub Copilot (CLI and Copilot Chat in VS Code) · pi-agent · omp · OpenClaw · zaly · Gemini CLI · Antigravity · Hermes · and CSV/JSONL logs of your own API requests.

Use H to switch tools live, or opentab --harness NAME to start with one.

What each tool's records support on top (recorded vs estimated cost, subagent tree, Turns, Trace, Tools, Context) is a per-tool capability table in the sources guide.

Capability table, data locations, setup, and the CSV/JSONL schema. Some tools need setup, such as Copilot CLI's opt-in telemetry export; opentab doctor helps you find the gaps.

Tip

Keep the history you want to explore. Claude Code and Gemini CLI can delete old local records automatically. OpenTab can only show what remains on disk. Check your retention settings.

About the money

Recorded spend comes from your tools' own records. A subscription session can record millions of tokens and $0.00 of per-token cost; your subscription is billed elsewhere.

API-equivalent estimates add list-price costs for that unpriced usage. This view is on by default and your choice is remembered; $ toggles it. Model rates ship in a bundled models.dev catalog, so ordinary browsing works offline. Refresh them when you choose with r in the P price table or opentab --refresh-models.

Estimates help you compare usage. Your provider's invoice may include subscriptions, discounts, credits, or markups that local records cannot tell OpenTab about. Pricing explained.

FAQ

Does any of this leave my machine?

Local browsing needs no cloud service or account and sends no telemetry. Your agents' files are opened read-only; OpenTab writes its own preferences, notes, caches, and the exports you request.

You control transfers: remote pulls fetch summaries from machines you choose, a price refresh fetches the models.dev catalog, and exports go wherever you put them. The live web browser binds to localhost by default. Privacy details.

Can I show it without exposing my project names and prompts?

Use opentab --demo, or D in-app, to anonymize titles, paths, prompt text, and absolute spend/token numbers in memory. The shape of the data stays real. Demo mode never writes those changes back to your history or saves your browsing state.

opentab --demo --html demo.html creates an anonymized report. You can also choose which categories to scramble; demo mode details.

How do I keep a longer history?

Claude Code deletes local transcripts after 30 days by default, and Gemini CLI's default retention removes chat recordings older than 30 days on launch. Once the source record disappears, its usage disappears from OpenTab too. OpenTab warns about these settings; opentab doctor reports them.

For long history:

• Claude Code: add "cleanupPeriodDays": 3650 to $CLAUDE_CONFIG_DIR/settings.json (default ~/.claude/settings.json).
• Gemini CLI: set "general": {"sessionRetention": {"enabled": false}} in ~/.gemini/settings.json, merging it with any existing general settings.

My tool isn't supported. Can I still use OpenTab?

If it can log API requests, point OpenTab at a CSV or JSONL of them. Turns, Tools, and Context become available when the log carries the relevant optional fields. Request log schema.

For another native integration, open an issue describing the tool and its record format.

Something looks off?

Run opentab doctor. It reports which tools it found, why others are missing, and what to fix, along with terminal support and the price catalog. It diagnoses without repairing or reading transcript content.

Full documentation · Troubleshooting · Windows & WSL

Development

CI runs Ruff, unit tests, and ShellCheck. CONTRIBUTING.md covers setup, checks, and commit conventions; Architecture explains the code.

License

MIT — see LICENSE.