Features

Providers & Routing

9 AI backends behind a unified interface. Full model tables and setup guides in the Providers documentation.

• -Provider lock per session — the provider chosen on the first turn sticks for the life of that chat tab.
• -Hot-swap provider or model mid-conversation — click a different chip, transcript replays automatically.
• -Availability probing — Ollama chip probes with a 2-second timeout so the chip reflects live status.
• -Preflight mismatch detection — pasting wrong key types triggers clear error messages.

Chat Experience

• -Multi-tab chats — every conversation is a real VS Code editor tab. Drag, split, pin, close independently.
• -Three entry points to start a chat: sidebar button, editor title bar icon, and status bar entry.
• -Streaming responses with a three-dot indicator and an immediate stop button.
• -Up/Down arrow cycles through previous prompts (terminal-style). Draft is preserved.
• -Enter sends, Shift+Enter inserts a newline. IME composition respected.
• -Auto-scroll pauses when you scroll up. A 'Jump to latest' button appears for new content.
• -Per-message actions on hover: Copy, Edit & resend, Regenerate.

Markdown Rendering

• -Assistant messages render as real Markdown — fenced code blocks, inline code, bold, italic, headings, lists, blockquotes, links.
• -Code blocks have a one-click Copy button with a success checkmark.
• -Safe React-element renderer — no dangerouslySetInnerHTML, no XSS surface.

Agentic Loop (Tool Approvals)

Every tool the model wants to run goes through a typed approval flow with visual feedback.

• -Inline diff shows before/after at the line level with compressed unchanged runs.
• -Open side-by-side VS Code diff editor if you prefer.
• -Tool message rows collapse to a single line with click to expand.
• -Long tool outputs collapse past 24 lines with a 'Show more' toggle.
• -TodoWrite renders as a live checklist with in-progress and completed states.

Attachments

• -@file mentions in the composer with fuzzy case-insensitive workspace search.
• -@terminal mention attaches the active terminal's last-command output.
• -Paperclip button opens the native OS file picker — multi-select from any drive.
• -Shift+drop files from VS Code file explorer or OS.
• -Paste screenshots (Ctrl+V) — saved locally and auto-attached.
• -Click attachment chip to preview; x button to remove before sending.
• -Auto-attach editor selection setting — selected code gets inlined in the message.
• -Validation: binary files rejected, files over 10 MB blocked, files over 1 MB prompt confirmation.

Sessions, Checkpoints & Rewind

• -Persistent sessions saved locally — restorable from the sidebar.
• -Automatic per-turn checkpoints before any Edit/Write/MultiEdit.
• -Rewind via the rewind icon on any user message — Fork, Restore, or Both.
• -Past Conversations picker with date buckets and metadata.
• -Carry-over summary with /fork <summary> to seed new sessions.

Context Management

• -Auto-compact fires when context window crosses the threshold (default 80%).
• -Context replay on provider/model switch — recent turns prepended as transcript.
• -Adaptive replay size auto-scales to the target model's context window.
• -Live context readout below the composer — input/output tokens, context %, cost.

Scheduled Messages

Schedule a message to fire at a specific time — useful for long tasks, follow-up prompts, or timed reminders.

/schedule                    → guided flow
/schedule refactor the auth  → skips message prompt
/schedule list               → view pending schedules
/schedule cancel <id>        → cancel a schedule

If VS Code is closed when the scheduled time passes, a notification shows what was missed with Send Now / Reschedule / Delete options.

Voice (Read Aloud)

Every reply has a play button that reads it aloud. Code blocks are summarised (“Code block in TypeScript, 12 lines”) instead of read out. Links are read by their label and file paths by their file name; markdown symbols and emoji are skipped. Stop Reading Aloud in the command palette stops playback.

/autospeak on      → read each reply of THIS chat aloud when it finishes
/autospeak off     → stop auto-play in this chat (play buttons still work)
/autospeak status  → whether this chat auto-plays, and which voice is used

Voices: Deepgram Aura-2 with your own Deepgram API key, or the free system voice. Add the key in Settings → Speech to Text & Text to Speech → Deepgram settings. It is stored with your other credentials, never in settings.json. Deepgram bills per character.

Voice Input (Speech to Text)

Push-to-talk: click the microphone in the chat composer (or press Ctrl+Alt+V, ⌘⌥V on macOS), speak, and click again. The words are inserted where the cursor is, for you to review before sending. Esc cancels. Speech to text uses Deepgram Nova-3 with your own Deepgram API key.

Wake word:say “hello coder” followed by your request and it is sent hands-free — or say “hello coder” alone, wait for the chime, then the request. A request spoken while a reply is still being written is added to that turn, like a typed one.

/hello-coder on      → turn the wake word on and start listening in this window
/hello-coder off     → turn it off and close the microphone
/hello-coder status  → whether it is on, and whether this window is listening

Privacy and cost: listening never starts by itself — after a reload, press Start listening in the composer. While listening, speech is detected on your machine and only what you say is sent to Deepgram (billed by audio length); replies being read aloud are ignored. The status bar shows when the microphone is listening. Desktop VS Code records with the system microphone (built in on Windows; SoX on macOS; arecord, SoX or ffmpeg on Linux). In web access the browser records — this needs an HTTPS address and your permission.

Cost & Usage

Every turn reports token usage — input, output, and cache reads/writes — plus an estimated cost for the turn and a running total for the session. Cache reads are shown separately because they are billed at a fraction of fresh input.

/cost    → this session's tokens and estimated cost
/usage   → MR Coder token usage (rolling window, weekly, monthly)

The context meter in the chat header shows how much of the model's window is in use, so you can see compaction coming before it happens.

Web Remote Access & Tunnel

Open your workspace from any browser. Opt-in via mrMultiCoder.webAccess.enabled.

• - Two-port design: an auth proxy in front, the VS Code Server internal-only behind it
• - JWT session cookies, bcrypt-hashed passwords, brute-force rate limiting
• - Per-user project permissions — restrict a user to specific folders
• - Extension allowlist enforced by VS Code 1.96+
• - Optional Cloudflare tunnel for a public domain

Full setup and the security model in the Web Remote Access guide.

MCP Servers (Model Context Protocol)

Connect external tools and endpoints from a visual modal — type /mcp in any chat, or use the MCP button in the composer. API keys go to your OS keychain, never to settings.json.

Jira scopes per question, not per connection. There is no project picker — Jira exposes no project-scoping header, so you narrow in the chat instead: “list my Jira projects” → “epics in PROJ” → “what's under PROJ-42?” to reach its stories and tasks. You only ever see what your Atlassian account can already see.

Servers are stored in mrMultiCoder.mcpServers — the same shape as Claude Code's config. Chat-capable endpoints (Anthropic/OpenAI Compatible) appear in the provider chip; tool servers stay in the tool plumbing.

Project Conventions (CODER.md)

Run /init to launch a wizard that asks seven questions and writes a single CODER.md at the repo root. It gets wired into every provider's memory file automatically.

• -Max lines per file, tech stack, testing policy, comment style, commit style, dependency policy, off-limits paths.
• -Idempotent — re-running /init never duplicates the reference block.
• -CODER.md is loaded first in the auto-memory chain across all providers.

Security & Workspace Trust

• -API keys live only in your OS secure keychain — never in configuration files, never synced, never logged.
• -Workspace trust respected end-to-end — edit, write, and bash tools refuse to run in untrusted workspaces.
• -CLI providers are never bundled — auto-detected on PATH with install hints.
• -Proposed-content store is session-scoped and automatically cleaned up.

Developer Analytics Dashboard

A full-featured, self-contained analytics dashboard for tracking productivity, cost, and reliability. Full specification in the Analytics documentation.

• -Productivity Score (0-100) with five weighted components.
• -12 KPI cards with period-over-period deltas.
• -Streak tracking, peak hours, efficiency trends, session quality.
• -Cost intelligence — daily trends, cost per prompt, cost by project/provider.
• -Model productivity ranking table.
• -Export to CSV or JSON. All data is local-only.

Settings Reference

All 97 settings, prefixed with mrMultiCoder. Open them with MR Multi coder: Open Settings or search the prefix in VS Code settings.

Keyboard Shortcuts

Full command reference with examples in the Commands documentation.