New
shift+↑ and shift+↓ scroll the conversation, next to the mouse wheel and pgup/pgdn.
Improved
- Scrolling with the mouse wheel could fill the message box with your earlier prompts instead of scrolling the conversation, in terminals that turn the wheel into arrow keys. The wheel no longer sends arrow keys while teveus is open, so ↑ and ↓ only browse your earlier prompts.
New
- teveus asks "Do you trust this folder?" the first time you open it in a folder, before anything is read or started. Answer yes once and it doesn't ask again there or in the folders inside it; answer no and it exits.
- Links in the conversation are clickable: cmd+click (ctrl+click on Windows and Linux) opens them in terminals that support links, including ones too long to fit on one line.
Improved
- Scrolling down could suddenly scroll the conversation up instead, after a mouse release the terminal never reported (for example when you cmd+clicked a link). A drag that lost its release now ends.
New
- Your own slash commands and skills on the Direct API engine. Markdown files in
.claude/commands/ (for example review.md becomes /review, with $ARGUMENTS and $1 filling in what you type after it) and in .claude/skills/, in the project or in ~/.claude, show up in the / menu. The model can also load a skill on its own when a task matches it.
/hooks: run your own commands before or after the model's tool calls, such as a formatter after each edit or a check that blocks a command. They are kept in teveus's settings folder, never in the project, so a repository can't run commands by itself.- Long-running commands in the background on the Direct API engine: the model can start a dev server or a watcher, read its output later and stop it. They all stop when you close the session.
- Notebooks (
.ipynb) on the Direct API engine: the model reads them cell by cell, without the bulky outputs, and can replace, insert or delete a cell.
/search text finds words in the open conversation (pick a match to jump to it) and in your earlier conversations in this folder (pick one to resume it)./worktree, or "New session in a git worktree" in /sessions: a session gets its own copy of the repository on its own branch, so several sessions can edit at once without overwriting each other. Closing it removes the copy unless it holds uncommitted work./subagent (or /settings → Research subagent model): run the research subagents on a cheaper model. Off until you choose one; they use the conversation's model by default, and fall back to it if the chosen one can't be reached.- What each session has used shows in
/sessions, with the total, and in the status bar when the sidebar is hidden.
Improved
- Long sessions on the Direct API engine use fewer tokens without losing anything: big tool output that is many steps old (a file read, a test log) is replaced in what the model is sent by a one-line note, and the model runs the tool again if it needs it. It happens in batches so the provider's prompt cache keeps working, and your transcript and
/resume still show the full output.
- Reading a file that hasn't changed since the model last read it now answers "unchanged, it's above" instead of sending the file again.
- When a conversation nears the model's context limit, old tool output is dropped first, and the conversation is only summarised if that isn't enough. That saves the cost of the summary.
- A session you aren't looking at now sends a desktop notification whenever it finishes or needs approval, naming the session, however short its turn was.
New
- Several conversations at once in one window.
/sessions lists them and lets you start, switch to or close one; alt+1 to alt+9 jump straight to a session. Sessions keep working while you look at another, and a line of tabs under the header names each one, with ● for one that is working and ! for one waiting for your approval. On macOS Terminal, turn on "Use Option as Meta Key" for the alt keys.
Improved
- Typing
/ now lists only the eleven commands most people need (/model, /login, /sessions, /mode, /resume, /undo, /diff, /clear, /settings, /help, /exit). Every other command, including Claude Code's own, still works and appears as soon as you type part of its name.
- Commits and pull requests made through teveus no longer carry a "Co-Authored-By" or "Generated with" line: not from Claude Code, and not from the Direct API engine. To bring them back, turn on "Credit AI in commits" in
/settings.
New
- teveus tells you when a new version is out and asks before updating, showing what's new. It checks GitHub at most once a day; choose "Later", skip that version, or turn checks off in
/settings. /update checks right away.
- Updating from inside teveus uses the way you installed it (the installer, with its checksum check, or
go install).
Improved
- The Claude Code notice and the update prompt are easier to scan, with a colour for each point and each choice.
- The installer's closing message now says to connect Claude Code with
/login.
Improved
- teveus now starts with nothing connected. Choose where the AI comes from with
/login: your installed Claude Code, or any provider with your own key.
New
/effort sets how hard the model thinks, from low to max, on both engines. Models that don't support it simply run at their default.- Settings with levels (effort, minimal code, guidance) use a coloured slider: move with ← →, and the whole box takes the colour of the level you pick.
/minimal asks the model to write only the code the task needs (off, lite, full or strict), and /trim reviews your changes for code you don't need./purge deletes saved conversations and prompt history in one step.
Improved
- The usage savers (short answers, lean tools, minimal code) are now off until you turn them on.
- When a provider is busy or rate-limited, teveus waits and tries again (up to three times) instead of stopping.
New
- Paste or drop images: drop a screenshot on the terminal or press
ctrl+v, and it's attached as [Image #1]. Large images are scaled down automatically.
shift+enter starts a new line (or \ then enter in terminals that can't tell the keys apart).- Long pastes fold into
[Pasted text #1 +40 lines] so the input stays tidy.
- For the Direct API engine:
- reading web pages, and web search with a Brave or Tavily key
- sub-agents that research in parallel
- MCP servers from
.mcp.json, started only after you approve them with /mcp
- "always allow" rules saved per project, managed with
/permissions
- long conversations compacted automatically before they outgrow the model
teveus -p "…" runs one prompt without the interface, for scripts (-json adds cost and token counts).- Faint lines between turns make long conversations easier to follow.
Improved
- Long conversations scroll and stream much faster.
New
- First release: a friendly terminal for AI coding with two engines behind one interface.
- Claude Code: runs your installed Claude Code.
- Direct API: your own keys for OpenAI, Anthropic, Google Gemini, OpenRouter (with browser login), Groq, DeepSeek, xAI, Mistral, any OpenAI-compatible service, or local models with Ollama and LM Studio.
- Every change is shown as a coloured diff and waits for your approval. Four plain-language modes: ask first, auto-edit, plan only, autopilot.
- Command palette (
ctrl+k), @ to mention files, prompt history, /undo, /resume, /diff and /export.
- Sidebar with cost, context size, plan progress and usage limits.
- Six themes including high contrast, three guidance levels, reduce motion, and full keyboard use.
- A three-step first-run setup.