Changelog

This page lists what changed for you in recent Tokenade releases, newest first. Internal changes are left out. Tokenade updates itself (see Upgrade and uninstall); run tokenade --version to see which release you have.

1.2.2

Help pages match what the commands do. tokenade --help and each command's --help were checked against the code: install points to tokenade login; style lists auto and says the setting is global; llm-proxy covers serving without --upstream, --enable, and when install sets the proxy up for you; statusline and watch have their own help; codex-trust, ensure-hooks and audit-gains are listed; hidden flags are now shown (audit-claude-md --apply/--compress/--stale, gain --recent, search --html, index --json, upgrade --check, uninstall --keep-data); and the main help points to your own TOML compactors (add-compactor).
Plans and prices are current everywhere in the CLI. The quota message shows Pro at €19.90/month incl. tax in Europe and $24.90/month excl. tax elsewhere, and pay-as-you-save at 10% of the API-cost value saved. login and activate no longer mention machine limits on Free and Pro, which have none.
uninstall --dry-run only says what it would do. A few lines read "removed" although nothing was touched.

1.2.1

Codex is covered even if you never approve its hooks. The local LLM proxy now does for Codex what the hooks would do: folded tool output, scrubbed credentials, the style note and its savings figures. If the hooks do run, nothing is done twice.
The LLM proxy keeps Codex signed in with your ChatGPT account, instead of expecting an OpenAI API key that ChatGPT-account users do not have.
The LLM proxy steps in exactly where an agent's hooks are off. Tokenade reads whether each agent's hooks actually run (Codex approval, disableAllHooks, plugin enable lists) and only covers the gap.
The LLM proxy stops re-sending identical tool output. A second copy of an unchanged file or command output becomes a short pointer to the first, where the agent's own hooks do not already deduplicate.
Claude Desktop's Code tab gets the full set of Claude Code hooks, including web-search and MCP compaction, even without a separate claude command installed.
tokenade health says when Claude Code ignores hooks in a folder whose trust prompt was not accepted.
Web search can be left entirely to your agent. With TOKENADE_NO_WEBSEARCH=1, tokenade search now refuses to run and the rules Tokenade writes leave it out.
MCP savings in desktop apps are credited to the app (Claude Desktop, Codex app, Cursor) instead of "unknown".
mcp-wrap leaves the Codex app's own MCP server alone.
Tokenade never asks for administrator rights. Commands that would need sudo are printed for you to run.
The LLM proxy starts at boot for every agent on Windows and macOS, not only the last one set up; on Windows its savings are credited to the right agent and an update really restarts it. llm-proxy install --autostart works on an agent that is already set up.

1.2.0

mcp-wrap finds Claude Desktop installed from the Microsoft Store.
Many LLM proxy fixes: Qwen Code models with their own provider address, requests relayed with their own HTTP method, Gemini streaming, Codex tool output in every shape, Cursor-style editors (only command output is folded, never your prompt or files), a credential never let through when folding is skipped, plan limits respected, and the proxy's savings now recorded in your figures.
Savings figures are more honest. Context Tokenade adds is charged on every agent and every request that carries it; folded output is credited conservatively per re-read; what you recover with expand-ref is deducted; subagents are counted in their own conversation; dollar savings are priced in the tokens Claude actually bills. See How savings are measured.
Dashboards say what each figure counts, tokenade dashboard --json gives the realistic net, and tokenade dashboard says where the page went.
Links in fetched pages can be followed: tokenade web resolves relative links against the page.
More commands are summarised: flutter analyze, dart analyze, dotnet build, dotnet test, Vite builds, GoogleTest binaries, codex exec and codex review, Worktrunk wt merge, and on Windows Select-String, French dir, recursive Get-ChildItem and PowerShell cmdlets.
Commands are recognised through their wrappers: time cmd, bash -c "…", subshells ( … ), and a leading pwd, date, mkdir, touch, source or disown.

1.1.23

Many more documents read faithfully by tokenade read: Word (.docx, .doc), OpenDocument, RTF, EPUB, PowerPoint (.pptx, .ppt), FictionBook, PDF and Excel (.xls, .xlsx, .ods). Numbered lists, footnotes, links, text boxes, formulas, tables and headings keep their structure; hidden or deleted text is no longer served.
New formats: e-mails (.eml), Outlook messages (.msg), mailboxes (mbox), calendars (.ics), contact files (.vcf), browser network captures (.har) and Jupyter notebooks.
OCR and transcripts: tokenade read on an image returns the text it shows, a scanned PDF is read by OCR (with a local Tesseract), and automatic subtitles give a clean transcript. See Read documents and media.
Large CSV files are summarised by column, like the same table in a workbook.
Your permission rules apply to what Tokenade handles: Claude Code deny and ask rules for Read(…), WebFetch(domain:…) and commands now also hold for files, pages and commands Tokenade serves or wraps.
Sharper compactors for pytest, go test -json, journalctl, ip addr, ss, git log --stat / git diff --stat, cargo tree -d, npm ls, pacman, node --test, REST API lists and JSON records.
tokenade dashboard opens the page in your browser reliably.

1.1.22

tokenade uninstall no longer reaches above your project, and a copy of Tokenade that is not the installed one no longer touches your Claude Code settings.
Tokenade rules no longer pile up in your global CLAUDE.md.
search fixes: batch searches no longer return "no results" when other engines answered, highlighted words are no longer cut, and the same page is no longer counted twice.
read --prompt returns passages in document order and never a lone heading.
Listings that fit are delivered whole: git log, git ls-files, ps, pip list, ls -l, getent, gh lists, CSV/TSV and any output requested as JSON (--json, -o json).
git grep is compacted like grep, and a large grep names every file where it found something.

1.1.21

Install choices survive updates. --no-hook, --only, --per-project, --no-debug-log, --debug-log, --no-autocompact, --no-wrap-mcps, --profile and --with-* are remembered by every later install, upgrade and self-repair. tokenade install --defaults forgets them.
A custom multi-line status line stays whole; the Tokenade ticker goes on the last line.
A status line you removed is no longer put back on the next session (Claude Code, Qwen Code, Copilot CLI, Cursor CLI). tokenade statusline on restores it.

1.1.20

Long sessions are no longer overcharged for the context Tokenade added before a compaction.
Each saving is weighted with its own model's prices, in the terminal, the status line and the web dashboard.

1.1.19

Gemini CLI and Reasonix are no longer supported; Windsurf becomes Devin Desktop. --only windsurf remains an alias, and tokenade uninstall still removes their old hooks and settings.
Rules are written once per agent, in the global file each agent really loads, instead of being copied into every repository. Duplicates in Kilo, Copilot CLI and project CLAUDE.md files are removed.
Updates also refresh the rules files, on every invocation, even with auto-update off.
Uninstalling RTK also removes its leftover instructions.

1.1.18

Grok no longer loads Tokenade's rules twice.
MCP screenshots are no longer folded when the tool is not named "screenshot".
A git diff with many regenerated files shows the real change instead of only the first twenty files.

1.1.17

An update now restarts the running LLM proxies, which kept serving the previous version.