Chat & MCP

The right-hand chat panel is a Claude session anchored to your active vault. The model reads and writes vault files (notes, connections, environments) through a small set of MCP tools — same permission rules apply whether the model invokes them or you click the equivalent UI control.

Open the panel

Cmd+B toggles the chat panel. On first open the sidecar (Node process bundled with the app) spawns and connects to Anthropic’s API; subsequent toggles reuse the running process.

The panel header shows the current working directory the chat is anchored to — defaults to your active vault path; click it to point the session at a different folder.

What the model can do

Out of the box the chat has access to four MCP tool families that the bundled httui-mcp server exposes:

Family	Tools
notes	list_notes, read_note, create_note, update_note, search_notes
connections	list_connections, get_connection
environments	list_environments, get_environment_variables
blocks	list_blocks_in_note, get_block_settings, get_block_history

Notes / connections / environments tools route through the same file-backed stores the UI uses — connections.toml, envs/*.toml, runbooks/*.md. Blocks tools surface metadata only (settings, history) — no body content, no execution.

Permission system

Tool calls don’t run silently. Each one checks against this cascade before executing:

1. Bash → always asks the user (no auto-allow).
2. Edit / Write outside the session cwd → hard deny.
3. Read / Glob / Grep inside the session cwd → auto-allow.
4. Persisted rule (tool_permissions table, scope always) → apply.
5. Session rule (scope session) → apply.
6. Fallback → ask via the in-app permission banner.

The banner offers three scopes when you approve:

• Once — applies to this single call.
• Session — applies for the rest of this chat session.
• Always — persists in the local SQLite. Manage saved rules via the gear icon in the chat header.

For update_note the banner shows a compact diff with the file path and +N -M line stats. Clicking “View Diff” opens a side-by-side merge view (CodeMirror); the chat won’t apply the edit until you Allow.

Image attachments

Up to 20 images per message, 5MB each. Three input paths:

• File picker (paperclip icon).
• Clipboard paste — Cmd+V over the chat input.
• Tauri native drag-drop — drag an image from the OS file manager into the chat input region.

Images are normalized before send: side > 2048px scales down via Lanczos3, then re-encoded to JPEG Q85. Source files stay untouched on disk.

Wikilinks in your message

Type [[note-name]] in the chat input. httui scans the active vault for a matching note (case-insensitive stem match), inlines its content as a context block for the model, and keeps the literal [[...]] in the message log so future scrolls show what you originally wrote.

Sessions

The chat stores sessions in the local SQLite (notes.db). Session list lives in the left tab of the chat panel. Each session keeps its own:

• Message history
• CWD anchor
• Persisted permission rules
• Token-usage roll-up

The Anthropic-side claude_session_id round-trips so a session resumes intact after an app restart. If the resume fails (rare — backend changes, deleted session id), the panel offers “Continue as a new conversation” — same messages locally, fresh claude_session_id upstream.

Custom MCP tools

The bundled httui-mcp server is just one MCP entry. Extra servers configure under ~/.config/httui/user.toml:

[mcp.servers.my-server]
command = "/usr/local/bin/my-mcp"
args = ["--vault", "${VAULT_PATH}"]

Each entry spawns a local process the chat can route tool calls to alongside the built-ins. The standard MCP stdio protocol applies; see the MCP spec for the surface a server has to expose.

Token usage

Per-day + per-session token roll-ups live under the Usage tab. Tracks input_tokens / output_tokens / cache_read_tokens separately so you can see the cache hit rate (cache reads are 10× cheaper than fresh reads). Cache efficiency lands as a single percentage at the top.

Things to know

• The chat is Anthropic-only today. No swap-in of OpenAI / open models — the protocol assumptions in the sidecar are Claude-specific.
• The chat is online-only at the moment. No offline mode.
• The chat respects the same conflict banner as the editor — if it tries to write a note that’s been edited externally during the chat session, the conflict banner fires before the diff and you decide whether to merge.