Features

Everything Codetta puts in front of you, in detail. AI panel, Claude Code, SFTP, MCP, multi-workspace — and the small choices that make them feel right.

Agent Mode · v0.5.0

An editor that becomes an agent

Flip the whole window into a dedicated agent workspace — a slim workspace rail, a sessions list, and a conversation-first chat as the main surface. Toggle with the title-bar Agent button or Ctrl+Shift+A; your sessions, changes, and tasks come with you.

  • Conversation-first chat — tool calls collapse to glanceable chips: a quiet "read 16 files", a loud "edited run.ts ×3". Click any chip to expand the diff or output.
  • Changes & tasks at a glance — edits roll into a Changes review card; the agent's to-do list lives in a sidebar section, not buried in the transcript.
  • Plugins & customizations — install Claude Code plugins from any GitHub marketplace; manage Instructions, Skills, MCP, tool access, providers, and privacy in one tabbed panel.
  • Plan-mode approval — review the agent's plan in full, then Approve & start, or keep refining.
# The agent's actions, as glanceable chips You: refactor the enrollment sweep Now hook into contacts/actions.ts: [🔍 read 6 files] [✦ edited actions.ts ×3] Let me typecheck: [➜ ran tsc → clean] ┌ Changes · 4 files · +120 −38 ──[Review]┐ └─────────────────────────────────────────┘
AI Panel

Bring your own model

One panel. Anthropic API, OpenAI API, Ollama (local), or Claude Code. Pick per-chat, switch mid-conversation, and pay nobody for the editor itself — your keys, your spend.

  • Claude Code uses your existing Claude Pro / Max subscription with no per-token bill.
  • Anthropic + OpenAI use your API key locally; nothing routes through us.
  • Ollama runs entirely on your machine — qwen, llama, deepseek, whatever you've pulled.
  • Per-chat tabs with provider badges (CC / Cl / AI / OL), drag-to-reorder, refresh-resume mid-stream.
Codetta's model browser modal showing All providers / Local (Ollama) / Reasoning / General-purpose / Coder tabs, with installed Ollama models (qwen3, gpt-oss, deepseek-r1, codellama) and an Install button per model
Claude Code

The Claude CLI, but with a UI

Codetta wraps the official claude CLI as a first-class provider. Subscribe once on Anthropic.com; spend zero dollars in Codetta. Permission cards, tool-by-tool always-allow rules, refresh-resume mid-stream, and per-chat session pinning so the CLI's context cache stays warm.

  • Three-tier always-allow — Allow this session, Always allow this tool, Always allow this filetype.
  • WebFetch + WebSearch gated through the same overlay so you can grant network access deliberately.
  • Refresh-resume: reload the page mid-stream, the chat re-attaches and replays buffered events.
  • Per-tool running rows with code preview — see what Edit / Write / Bash is about to do before the agent commits.
# Permission card surfaces in the editor PreToolUse: Edit src/store.ts └─ old_string: "setActiveWorkspace: async ..." └─ new_string: "setActiveWorkspace: async (id) => {" [✕ Deny] [✓ Allow this session] [✓✓ Always allow Edit on .ts] [✓✓ Always allow Edit] [✓ Allow once]
SFTP · v0.4.0

Edit remote like local

Connect to any SSH server (password or key), browse the remote tree in a sidebar, and edit files in the same Monaco buffer you use for local code. Auto-push on save, recursive directory sync, and a connection pool so browsing doesn't reconnect on every click.

  • Connection pooling — one persistent SSH channel per profile; a list-dir click drops from ~1.5s to ~50ms.
  • SSH key auth — optional private-key path on every profile, falls back to password.
  • Auto-push on save — opt in per file with a click; tab title shows ↥ for opted-in buffers.
  • Recursive sync — right-click a folder to upload or download whole trees over a single session.
  • Type-to-confirm delete — folder deletes require typing the name back. No stray-click disasters.
// Save → auto-push, fire-and-forget await fs.writeFile(path, content); const link = lookupRemoteLink(wsId, path); if (link?.autoPush) { const session = getActiveSftp(wsId); if (session?.profileId === link.profileId) { invoke("sftp_write_file", { args: { ...session.conn, path: link.remotePath, content }, }).then(() => toast("↥ Auto-pushed")); } }
Workspaces

Many projects, one process

One running editor instance, every project as a tab in the activity bar. Per-workspace state lives in %AppData%/codetta/workspaces/{id}/state.json with atomic writes, so switching projects is instant and safe.

  • Pop-out terminals — drag any terminal into a separate OS window; PTY survives the move.
  • Split panes — drag tabs to any edge to split horizontally or vertically.
  • AI chat tabs live in editor panes alongside files; pop the chat into its own pane while you read the diff.
  • Refresh-resume survives across the whole UI — close the laptop, open it tomorrow, your panes are exactly as you left them.
Codetta showing a Markdown source file in the left editor pane and its live rendered preview in the right pane — drag-to-split layout in action
MCP

Model Context Protocol, built in

Browse, install, and configure MCP servers without leaving the editor. Filesystem, Git, GitHub, Postgres, Puppeteer — pre-configured catalog with placeholder prompts so you can't ship a broken config to ~/.claude.json.

  • One-click install per server with input validation (path checks, GitHub-token prefixes, postgres URL schemes).
  • Per-scope toggle — install user-level or per-project, with badges so you can tell at a glance.
  • Editable post-install — re-run install to update env vars or paths without manually editing JSON.
// MCP server browser → Settings 📁 filesystem [↻ User] [+ Project] ⎇ git [+ User] [↻ Project] 🐙 github [+ User] [+ Project] // needs token 🐘 postgres [+ User] [+ Project] // needs URL 🎭 puppeteer [+ User] [+ Project] 🗄 sqlite [+ User] [+ Project]
Trust

Per-tool, per-prefix, per-extension

Three-tier always-allow lets the agent move fast on the boring stuff while still showing you the card on anything material. Bash gets per-prefix rules so grep + npm auto-allow but rm still asks.

  • Allow this session — in-memory only, resets on app restart.
  • Always allow <tool> — persisted; manage in Settings.
  • Always allow Bash <prefix> — persisted per first-token (grep, npm, git diff).
  • Always allow <tool> on .ext — persisted per filetype (Edit on .ts, Read on .json).
// localStorage allowlist (Settings can edit) [ "Read", // blanket tool "Bash:grep", "Bash:npm", // prefix gates "Ext:.ts:Edit", // per-filetype "Ext:.json:Read" ]