teveus Download

teveus docs

Everything teveus can do, on one page. Press ? inside the app for the short version.

Install

On macOS and Linux, paste this into a terminal.

  • Go: go install github.com/stawan15/teveus@latest
  • Windows or by hand: download an archive from Releases.
  • Installer options: set TEVEUS_INSTALL_DIR=/usr/local/bin or TEVEUS_VERSION=v0.1.1 before sh. The installer checks the download against the release's SHA-256 checksums.

teveus checks GitHub for a new version at most once a day and asks before updating, showing what's new. Choose "Later", skip that version, or turn checks off in /settings. /update checks right away and updates the way you installed (the installer or go install). Running the install line again also updates.

To uninstall, delete the program and, if you want, your settings and saved keys:

rm ~/.local/bin/teveus
rm -rf ~/.config/teveus

Get started

  1. Open a terminal in your project folder and run teveus.
  2. Connect an AI with /login. Nothing is connected until you choose.
  3. Type what you want in plain words, like "find the bug in the login page and fix it", and press enter.

teveus shows every change as a coloured diff before it happens. Press y to allow it, a to always allow it in this project, or n to say no.

The first time, a three-question setup picks how much guidance you want (guided, standard or pro), a colour theme, and how to connect. Run /setup to do it again.

Connect an AI

teveus has two engines behind one interface. Switch with /engine.

Claude Code

For Claude models on your Claude Pro or Max plan. Install Claude Code from Anthropic and sign in to it; teveus runs it as it is. You sign in through Anthropic's own login, and teveus never sees your Claude credentials. The first time, teveus asks you to confirm that it should run your Claude Code under Anthropic's terms.

Direct API

Your own keys, billed by the provider per token. Run /login and pick one:

OpenRouterLog in with your browser, no key to copy. Hundreds of models; type free in /model to find free ones.
OpenAI, Anthropic, Google Gemini, Groq, DeepSeek, xAI, MistralPaste an API key. It's checked before it's saved.
Ollama, LM StudioRun on your own machine. Nothing to set up if the app is running.
CustomAny OpenAI-compatible API: give its base URL (e.g. https://api.together.xyz/v1) and a key.

Keys are stored only on your computer, in ~/.config/teveus/auth.json, readable only by you. Undo, web tools, MCP, sub-agents and saved permission rules work on this engine; Claude Code brings its own.

Modes

The mode decides how much teveus asks before acting. It's always shown on screen; press shift+tab to cycle through them.

Mode-modeWhat it does
ask firstdefaultAsks before every edit, command, web request and MCP tool. Start here.
auto-editacceptEditsEdits files inside the project without asking. Commands and anything outside the project still ask.
plan onlyplanReads and plans, changes nothing. Asks before using the web.
autopilotautoApproves everything except your deny rules. Use it only in projects you trust.

Keys

enterSend your message
shift+enterNew line (or ctrl+j, or \ then enter)
ctrl+kCommand palette: find any action
/Slash commands
@Attach a file by name
ctrl+vPaste an image from the clipboard
↑ ↓Prompt history
shift+tabSwitch mode
esc escStop the current task
ctrl+oExpand or collapse tool output
ctrl+bShow or hide the sidebar
ctrl+yCopy the last reply
dragSelect text; it's copied when you let go. Double-click a word, triple-click a line.
pgup pgdnScroll the conversation
?Show every shortcut
ctrl+c ctrl+cQuit

When a change is waiting for you: y allow · a always allow · n deny.

Drop a screenshot on the terminal or press ctrl+v to attach it as [Image #1]. Long pastes fold into [Pasted text #1 +40 lines].

Commands

Type / to see them all, or find them in the ctrl+k palette. On the Claude Code engine, Claude Code's own commands are listed too.

/login Connect Claude Code or a provider; /login brave or /login tavily adds a web search key
/logout Remove a saved provider key
/engine Switch engine: claude (Claude Code) or api (your own keys)
/model Switch model; search by name, or type free also /m
/mode Choose the permission mode (same as shift+tab)
/effort How hard the model thinks: auto, low, medium, high, xhigh, max
/clear Start a new conversation also /new /reset
/resume Reopen an earlier conversation in this folder also /r /continue /sessions
/undo Put back the files the last turn changed (Direct API) also /u
/diff Show what changed in the working tree (git)
/init Write AGENTS.md / CLAUDE.md describing this project
/export Save this conversation as Markdown
/copy Copy the last response
/concise Short, direct answers on or off
/lean Send fewer tool definitions (about 2.2k tokens less per request)
/minimal How much code the model may write: off, lite, full, strict
/trim Review uncommitted changes for code you don't need; /trim all checks the whole project
/permissions List, add and remove saved allow/deny rules
/mcp List MCP servers and approve a project's .mcp.json
/purge Delete saved conversations and prompt history
/update Check for a new version of teveus and install it
/theme Change the colour theme
/sidebar Show or hide the sidebar
/details Expand or collapse tool output
/mouse Hand the mouse back to your terminal for its own selection
/settings All preferences in one place also /prefs /config
/setup Run the first-time setup again
/help Keyboard shortcuts and commands also /? /h
/exit Quit also /q /quit
!command Run a shell command, e.g. !npm test

Effort and savers

/effort sets how hard the model thinks, on a slider you move with ← →:

  • auto: the model's own default
  • low: quickest and cheapest, for questions and small edits
  • medium: everyday coding
  • high: tricky bugs and design decisions
  • xhigh: deep reasoning, the sweet spot on the newest models
  • max: the hardest problems; costs the most

Models that don't support effort simply run at their default.

Three savers cut usage. They're off until you turn them on, and -full turns them off for one run:

  • /concise: short, direct answers.
  • /lean: leaves rarely used tools out of each request.
  • /minimal: off, lite, full or strict. The model reuses what exists and writes only what the task needs.

/trim reviews your uncommitted changes for code you don't need; /trim all checks the whole project.

Web, MCP and helpers

On the Direct API engine, the model can also:

  • Read web pages (WebFetch), after you approve each one.
  • Search the web (WebSearch) with a Brave Search or Tavily key: /login brave, /login tavily, or the BRAVE_API_KEY / TAVILY_API_KEY environment variables.
  • Send research to sub-agents that read in parallel and can't change anything.
  • Use MCP servers from a project's .mcp.json or your own ~/.config/teveus/mcp.json.

MCP files use the same format as Claude Code. Servers run as a local program (stdio) or at an address ("type": "http"):

{
  "mcpServers": {
    "tracker": {
      "type": "http",
      "url": "https://mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer ${TRACKER_TOKEN}" }
    },
    "files": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "."] }
  }
}

${NAME} in a URL or header is read from your environment, so tokens stay out of the file.

A project's .mcp.json can start programs, so its servers stay off until you approve that exact file with /mcp. Editing the file withdraws the approval.

Long conversations are compacted automatically before they outgrow the model.

Permission rules

Answering a (always allow) saves a rule for this project. Manage rules with /permissions, add one with /permissions allow … or /permissions deny …. They use Claude Code's syntax:

BashAny command
Bash(npm test)Exactly this command
Bash(go test:*)Simple commands starting with go test
WebFetch(domain:go.dev)Fetching pages from one site
mcp__githubEvery tool of an MCP server; mcp__github__create_issue for one tool
Edit, Write, WebSearchA whole tool

Deny wins over allow and applies in every mode, even autopilot. Allow skips the question but never lifts plan mode. An allowed command prefix never covers a command chained with &&, ;, |, $( ) or redirections. Rules live in ~/.config/teveus/permissions.json, never in your repository. On the Claude Code engine, Claude Code's own rules apply.

Scripts and flags

Run one prompt without the interface and print the reply:

teveus -p "summarise what changed on this branch"
teveus -p "fix the failing test" -mode acceptEdits -json

Tools that need approval are refused unless -mode allows them. -json adds the cost and token counts.

-p "prompt"Run one prompt without the UI and print the reply
-jsonWith -p: print the reply, cost and token counts as JSON
-modedefault, acceptEdits, plan or auto
-engineclaude or api
-modelModel alias or ID, e.g. sonnet or google/gemini-3.5-flash
-cContinue the most recent conversation in this folder
-resume IDResume a conversation by ID
-themeclaude, tokyo-night, catppuccin, gruvbox, high-contrast, light
-fullTurn the usage savers off for this run
-claude PATHPath to the claude program
-versionPrint the version and exit

Settings and files

Change preferences with /settings. Everything is kept in ~/.config/teveus/ (%AppData%\teveus\ on Windows; set TEVEUS_CONFIG to move it), readable only by you.

FileContentsKept
auth.jsonAPI keys you saved with /loginUntil you /logout
settings.jsonYour preferences, and when the last update check ranAlways
history.jsonYour last 500 prompts, for the ↑ keyUntil /purge
sessions/Direct API conversations, for /resume30 days (keepSessionsDays; -1 keeps them)
permissions.jsonAllow/deny rules you savedUntil you remove them
mcp-approved.jsonWhich projects' MCP servers you approvedAlways
mcp.jsonYour own MCP servers, for every projectAlways

/purge deletes the saved conversations and prompt history at once. Claude Code keeps its own history in ~/.claude/.

Six themes: claude, tokyo-night, catppuccin, gruvbox, high-contrast and light. /settings also has the guidance level, reduce motion, sidebar and mouse options.

Privacy and security

teveus has no servers, no accounts and no telemetry. What you send goes only to the AI provider you picked, under their terms, and local models stay on your machine. The only request it makes on its own is the daily update check: the latest version number from GitHub, which you can turn off in /settings.

  • In the default mode, edits, commands, web requests and MCP tools all ask first.
  • The project folder is the boundary: reading or searching outside it asks first in every mode but autopilot.
  • teveus's file tools never touch credentials, in any mode: its own keys, Claude Code's login, SSH and GPG keys, and token files like .aws, .netrc and .npmrc.
  • Commands you approve run with your permissions. Keep your work in version control and use autopilot only in projects you trust.

The full details are in PRIVACY.md and SECURITY.md. Report vulnerabilities privately through GitHub.