Browser control. With "browser": {"enabled": true} in your personal config, the model can drive a real Chrome window: open pages, read a snapshot of their text, links and fields, click, type, pick options, press keys, switch tabs, run JavaScript and read the console. Opening a page asks per domain, like web_fetch. Chrome keeps its own profile, so sign-ins last between sessions. browser_screenshot lets the model see the page (the viewport, the full page or one element), optionally with each element's ref drawn on it. See Browser.
Images from tools. Tool results can now carry images to the model, on every provider: browser screenshots, and images returned by MCP tools (such as Playwright MCP's screenshots), which used to be replaced by a note.
Footer colors. The permission modes each have their own color (plan blue, accept edits violet, yolo red), effort is a magenta ramp with a bar that grows from low to max, and the footer now always shows whether the sandbox is on. Cost shows the session budget ($0.68/$2.00) and turns amber, then red, as it nears the cap; the context percentage matches its bar; subagents waiting for permission turn amber. See Footer colors.
A teal theme with a transparent background. The terminal UI now uses the site's teal accent in both dark and light, and no longer paints backgrounds of its own (footer chips, the task card, the composer's cursor line, Markdown headings and code): your terminal's background shows through. Diff lines keep their faint green and red tint.
@ mentions. Typing @ opens a file picker; each @path in a prompt attaches the file (with line numbers, so the model can edit it straight away), @path#L10-40 a range, @folder/ a listing, and @image.png an image for vision models. Pasted or dropped file paths become mentions, and mentions work in larik -p and larik serve too. See Composer.
! shell commands run straight from the prompt, in the sandbox, with their output sent along with your next prompt.
Prompt history per project: ↑/↓ to recall, ctrl+r to search. ctrl+g edits the prompt in $EDITOR.
A session picker:/sessions and /resume without an ID open a searchable list.
NVIDIA NIM as a provider preset: /connect nvidia-nim, or set NVIDIA_API_KEY. See Getting started.
An install script,scripts/install.sh, that builds and installs larik.
Task list. The model keeps a checklist with todo_write for work with several steps. The open items are pinned under the conversation during a turn, and /todos shows the list. See Commands and keys.
Leaving plan mode. In plan mode the model presents its plan with exit_plan_mode; approving it switches to the mode you pick and continues. See Permissions.
Custom commands from .claude/commands/*.md and ~/.claude/commands/, with $ARGUMENTS, $1… and !`cmd`. See Skills.
multi_edit makes several replacements in one file in a single call, all or none.
Image paste.ctrl+v attaches an image from the system clipboard, such as a screenshot.
/init writes an AGENTS.md for the project; /export saves the session as Markdown; /copy copies the last reply.
MCP resources, prompts and sign-in. Servers' resources can be attached with @server:uri and read by the model, prompts run as /mcp__server__prompt commands, and http and sse servers that need OAuth sign in with /mcp login. See MCP.
Prompt hooks ("type": "prompt") ask a model to decide instead of running a command. See Hooks.
LSP pull diagnostics and code actions. Servers that offer textDocument/diagnostic are asked for diagnostics after an edit, and the model can list and apply code actions, such as adding a missing import. See LSP.
A custom status line from a command of your own, compatible with Claude Code's statusLine scripts. See Status line.
Rebindable keys with the keybindings setting. See Rebinding keys.
Vim mode for the prompt: /vim, or Editor mode in /config. See Vim mode.
Mouse scrolling can be switched off in /config, so text can be selected without a modifier.
Debug mode and trace viewer.--debug, LARIK_DEBUG=1 or /debug on records every request as sent, each response with its timing, tokens and cost, the raw HTTP exchange, tool calls, permission answers and hooks. larik trace or /trace opens it in the browser: a timeline of each agent's requests and tools, a filterable list, and an inspector that diffs each prompt against the one before. --html exports one page. See Debug mode and traces.
The conversation scrolls above a fixed input row (Page Up/Down or the mouse wheel), and pickers and settings open just above the input with recent messages still visible.
Edit and file-write previews use syntax highlighting, and diffs show their changed lines as one highlighted block.
F2 and /info toggle a session sidebar that stays open while you type and while a turn runs. It is split into sections and shows only what is active: the task list, connected or failing MCP servers, running language servers and the skills used this session. See Keys and commands.
/model, /effort, /mode and shift+tab save your choice as the default for future launches. --model, --effort and --mode still override it for one launch.
Changing a provider's credentials takes effect straight away, without switching models.
Interrupting a turn at the wrong moment could leave a tool call without a result, which made every later request fail. Sessions already left that way are repaired on the next prompt.
An "always allow" rule for a command no longer matches the same command chained with others (git status; curl … | sh).
Sandboxed commands can no longer redirect the project's git metadata to run code the next time git runs outside the sandbox.
MCP servers using sse no longer disconnect right after connecting.
/compact can be canceled with esc, and prompts typed while it runs are queued instead of racing it. ctrl+c stops a running ! command.
Pasting into a picker filter or a /config text field no longer lands in the hidden prompt.
Changing routing mid-session no longer changes the tool definitions and invalidates the prompt cache.
Tool-call IDs from OpenAI-compatible, Ollama and Gemini models no longer repeat between turns.
ChatGPT plan sign-ins refresh reliably when several sessions share them.
Custom providers with type: "openai" or "anthropic" keep their own names for fallbacks and reasoning replay.
A shell command that leaves a background process running no longer hangs the turn.
Shared project settings can no longer re-enable a language server you disabled.
Notifications fire on terminals that don't report focus, and several smaller issues in the TUI, sessions and server were fixed.
Cheap and strong model routing. The main agent plans and reviews on a strong model while worker and explore subagents run on cheaper or local ones, across providers. /routing offers Balanced, Cheapest and Local-first presets, per-task tiers, fallbacks on rate limits and outages, and a session budget. See Model routing.
larik bench runs small self-checking coding tasks against several models and reports pass/fail, cost and time, so you can check a cheap model before trusting it with a role.
Routing shows what it saved, and cheap subagents can use a minimal prompt (role_options.<role>.context: "minimal").
Token saver filters recognized successful shell output before it enters model context. /config token_saver=true enables it; raw_output retrieves the exact private session copy without rerunning a command. See Tools and permissions.
Sandboxed bash can no longer write to the machine-wide /tmp. Each sandbox gets its own private scratch directory, exposed as $TMPDIR, so mktemp and build tools still work. See Sandbox.
Cross-platform process and file locking now allow the Windows binary to build. Windows shell commands require Bash and approval because Larik has no Windows sandbox.
The core harness: a Go + Bubble Tea terminal agent with Anthropic, OpenAI, Gemini and any OpenAI-compatible endpoint. See Getting started.
ChatGPT plan sign-in for Codex models, and native Ollama support that loads models with a usable context window and downloads models from the setup wizard.
Extensibility: an MCP client, Claude Code–compatible hooks, Agent Skills, and subagents through the task tool, including background subagents and git worktree isolation.
Language servers: diagnostics after every edit, plus an lsp navigation tool. See LSP.
OS sandbox for bash: Seatbelt on macOS, bubblewrap on Linux. See Sandbox.
A redesigned banner, footer and permission prompt; thinking collapses to a one-line summary.
Tool results show short paths.
Thinking time is saved with the session, so resumed history shows it.
/clear reloads the system prompt, undo history survives --resume, and compaction summaries point at the transcript file so the model can look up details.