What's new

The full history is in CHANGES.md.

0.19.0 — What fills a session’s context, and whether your plan fits

• See what fills a running session’s context. On Running sessions, click a session’s name. A grid like Claude Code’s /context shows how full the window is and what fills it: system prompt, tool definitions, tool calls and results, your prompts, Claude’s replies, skills and memory. Tables list the tools and skills in it, including tools that are loaded but never called, and how much the cache served. The total and the cache split are actual; the split by kind is estimated.
• Plan fit on the Scorecard. A new card says whether to stay on your plan or move up or down, and why. It shows your current limits, your usage pattern, each plan against your heaviest week and a busy 5-hour window, how much of the plan you leave unused, and what your usage would cost at API prices. It needs the Claude desktop app’s plan-limit history.
• Fixed: context size after a compaction. A running session that had compacted since its last reply kept showing its old, larger context. It now shows what compaction left.

0.18.1 — Running sessions work on Linux

• Fixed: running sessions on Linux. Running sessions didn’t show your Linux sessions, because Claude Code records a process’s start time differently there. They show up again, with their buttons.

0.18.0 — See Claude’s reply to each prompt

• Claude’s response in the prompt panel. Click any prompt (from Recent, a session or a list) and its details now show what Claude replied, right under your prompt. Thinking, tool calls and tool output are left out. If Claude only ran tools, or the transcript was deleted, it says so.
• Replies are not stored. The reply is read from the session’s transcript each time you open the prompt, so the dashboard’s database still holds only your prompts. Hide prompts hides the reply too.

0.17.1 — Running session buttons are back, close several at once

• Fixed: running sessions lost their buttons in 0.17.0. Every running session said “No matching process found”, and Interrupt, Compact, Close session, Force kill and Rename were missing. They are back.
• Close several sessions at once. On Running sessions, tick the idle ones (or Select all idle) and click Close selected. A session that is working is never closed this way.
• Fewer accidental closes. A working session is marked “Working now”. Closing it, or a session you used in the last 50 minutes, warns you before the second click.
• Recent, in the order you used things. Past sessions are sorted by their last prompt, and each session’s prompts list newest first. A sort button on each session flips its prompts, and the one at the top puts the last-used session on top or at the bottom. Running sessions always stay on top.
• Hide prompts in the side panel. The side panel has its own Show/Hide prompts button, and it keeps the panel open.

0.17.0 — Claude Code only, a link for every page

• Claude Code only. The dashboard no longer reads Codex, Gemini CLI or Cursor. The agent chips, the Agents page, the Cursor key and Cursor team spend on Billed vs local, and the OpenAI, Gemini and Cursor prices are gone. Claude desktop (Cowork) sessions are still read. Your next sync removes other agents’ sessions, so they no longer count in any total.
• A link for every page. Every page has its own address, and so does the side panel when it shows a session or a prompt. Refreshing keeps you where you were instead of going back to Context hygiene, and the browser’s Back and Forward buttons work.
• See which copy is running. If you run the dashboard both from npm and from a source checkout, the header now tells you which one this is.

0.16.0 — See and name your running sessions

• See which session is which. Each running session has a Prompts button. It shows the first prompt and the latest five, so you know what a session was doing before you close it.
• Name a session from the dashboard. Rename types /rename <name> into that session’s terminal, so Claude Code saves the name itself. The new name shows on the dashboard right away.
• Works in herdr. If you run Claude Code inside herdr, the dashboard can now reach those sessions. Compact and Rename type into the session’s herdr pane, Rename also renames it in herdr’s agents sidebar, Interrupt presses Esc there, and Hand over opens the new session as a herdr tab in the same workspace.

0.15.1 — Running sessions show up again

• Fixed: running sessions went missing on the 1st to the 9th of every month. On those days the dashboard misread when each session started, decided they had all ended, and hid them. Close, Force kill and Compact refused to work for the same reason. Both work on every day of the month again.

0.15.0 — Subagent models, numbers you can trust, history that outlives cleanup

• Subagent models. A new page shows each subagent type (Explore, general-purpose and your own agents), the models it ran on, and what a typical run cost, how many steps it took and how much it wrote. It suggests a cheaper model only when that model has enough runs of its own and really cost less per run, and it never shows a “you could save” figure. Your own agent files get the exact line to add.
• Try a model and see. Start experiment runs your subagents on one model for a while (it shows you the one line it adds to Claude Code’s settings first, and keeps a backup). The page then compares each subagent type before and after. Stop puts it back.
• Your own calendar days. Days, “today” and the billing period now follow your computer’s time zone, so late-night work no longer lands on the wrong day. Set CLAUDE_FINOPS_TZ to use another zone.
• Waste detection counts everything. It used to stop at 15 items per rule and show that as the count. It now counts all of them, says whether a long prompt came from you or from a script, and no longer calls a pipeline re-sending its template a “repeated prompt”.
• Forecast and Burn rate hold steady. They no longer change with the date range you pick, and days since your last sync are projected instead of counted as $0.
• Prices. Claude Sonnet 5.5 is priced. A new Claude model that isn’t in the price table yet is priced like the newest model of its family and marked, instead of showing $0. OpenAI and Gemini prices were checked against their official pages.
• Cursor shows only what it really recorded. Cursor’s agent transcripts have no times or token counts, so they are no longer read. Cursor IDE chats, which have both, still are.
• Fresh by itself, and history kept. The dashboard re-reads your sessions once its data is an hour old. When Claude Code deletes old transcripts, the sessions already in your dashboard stay.
• Many smaller corrections across the overview, Context hygiene, Usage, Model analysis, Prompt intelligence, Compare models and Skills & MCP. The full list is in CHANGES.md.

After upgrading, click Sync once so your history is re-read with local days and the new prices.

0.14.0 — Limits for one conversation, a simpler Budgets page, Jev

• A token limit for each Claude Code conversation. Set it on the Budgets page (block ③). It counts everything a conversation uses, including re-reading its context. A big project can have its own limit, or none. The Budgets page shows your largest conversation against its limit, and Sessions marks conversations near (amber) or over (red) theirs.
• Live warnings inside Claude Code, if you want them. Your limits work without this. Install it and Claude Code warns you as a conversation grows, and asks “continue?” before its next step once it reaches its limit. It can’t end a conversation, and if it ever fails it simply lets Claude carry on.
• Set a limit on a conversation that’s already running. On Running sessions, Set limit shows what that conversation has used so far and suggests a limit a bit above it. It applies from the next step.
• A Budgets page anyone can use. Three numbered blocks (money, tokens, one conversation), with the advanced settings folded away. Every box suggests values from your own last 30 days, accepts short numbers like 20M or $3,000, and tells you what’s wrong before you save. How does this work? opens a short picture guide.
• A Settings page for API keys. The optional keys for Billed vs local now live on their own page. You can add or remove them without the terminal, and the page only ever shows a key’s last four characters.
• Jev (fast decisions). A new page for TypeSafe’s Jev, a very fast, very cheap AI that only makes decisions. Install its Claude Code plugin in one click, add your Jev key, and see how much of your own work was really just a decision. Jev makes the apps you build cheaper, not Claude Code itself, and the page says so.
• Fixed: the Custom date range picker opens in the right place and closes on a second click.

0.13.0 — Flat icons everywhere, the dashboard opens itself

• Opens in your browser. Run claude-finops and the dashboard opens in your default browser as soon as it’s ready. There’s no address to copy. On a machine without a browser (over SSH, for example) nothing opens, and the printed address still works.
• One clean icon style. The emoji and text symbols in the sidebar, buttons, alerts and card headings are now flat line icons that match the top bar. They follow the light and dark theme and look the same on every computer. Severity markers are small flat dots in green, amber, orange and red.

0.12.0 — A cleaner top bar, colour-graded context metrics

• A cleaner top bar with more at a glance. Every page now shows estimated spend, sessions, prompts, requests and active days for your current filters, next to the page title. The date range and active filters sit on a small line under the title, and the filters take less room.
• Icons that say what they do. Tour, Recent, Hide prompts, theme and Reload are one tidy toolbar with clear icons: a history clock for Recent (still R), an eye for Hide prompts, a moon or sun for the theme. Sync has a database icon, so it no longer looks like Reload.
• Colour-graded context metrics. On Context hygiene, the cards turn green, amber, orange or red depending on how much of your spend sat at large context. Hover a card to see its bands.
• The tour explains every metric. The Context hygiene tour now has a step for each card: what it measures, why it matters and how to read it.
• A friendlier loader. While a page computes you see its outline, what it is doing right now, and a tip.

0.11.0 — Hide prompts, Recent panel

• ☰ Recent, from any page. A side panel opens from the top bar (or press R). Sessions running right now come first, with Resume, Compact and Stop. Stop interrupts the current turn, and Compact and Stop each need two clicks. Past sessions follow, newest first. Click one to see its prompts, and click a prompt for its full details. Search finds prompts by their text across all dates.
• Hide prompts. A top-bar button hides prompt text and session titles as dots everywhere in the dashboard, including tooltips, chart labels and the Recent panel. Use it before you share your screen. Your browser remembers the setting. Exports and the database are unchanged.

0.10.1 — Maintenance release

It contains the same dashboard as 0.10.0; only the version number changed. If you’re already on 0.10.0 there’s nothing new to see, and npx claude-finops picks up 0.10.1 automatically.

0.10.0 — Act now

• ⚡ Act now on the overview. A short list of one-click actions that apply to you right now: compact or hand over a running session carrying heavy context, turn a shell command you keep re-running into a skill, move an instruction you keep re-typing into CLAUDE.md, or install the statusline.
• Resume and Compact in every session row. The Sessions table and Cost rankings have a Resume button that copies claude --resume <id>, and running sessions also get a Compact button.
• “Focus on these first” is now “What to change”. Each item is labelled Fix once or Habit, shows the evidence behind it in one line, and can be hidden until that evidence changes.
• The statusline works. It now uses Claude Code’s documented settings format, handles install paths that contain spaces, and shows how much of your 5-hour limit you’ve used.

0.9.0 — Peak hours, plan-limit history, Claude desktop sessions

• Peak hours: a weekday × hour heatmap of your spend, in your local time.
• Plan limits over time: your 5-hour and weekly plan usage from the Claude desktop app, charted with your peaks.
• Claude desktop (Cowork) sessions now load alongside your other Claude sessions.
• Today against yesterday on the overview’s spend card.
• Claude 4.x and Haiku 3.5 are now priced at Anthropic’s published list rates.

After upgrading, click ⟳ Sync once (or start with --rebuild).

0.8.0 — Cost figures corrected

Requests were being counted more than once, two models had the wrong list price, and a long-context surcharge was applied where it didn’t belong. Totals fall sharply as a result: on the author’s own data, from about $17,000 to about $3,400. Several features that relied on assumptions the transcripts can’t support were removed, including the model-switch back-test and the letter grade. The warehouse rebuilds itself on first launch after upgrading.