Skip to content

AI Chat

AI Chat is the heart of Blyck. It embeds the Claude Code CLI (Anthropic) or the Codex CLI directly inside the app, calling models with your own account and credits. Conversations never pass through Blyck servers and are stored only locally (userData/chats/).

The first time you open the AI Chat panel, the AI Readiness Wizard runs automatically.

  1. CLI detection — Checks whether the Claude Code CLI / Codex CLI is installed on your system.
  2. Install guidance — If not installed, it shows the install command and links.
  3. Login check — Verifies that claude login / codex login has been completed.

Once you pass the wizard you can start chatting right away. The wizard won’t appear again afterward.

MethodDescription
Ctrl+Shift+ASplit the active panel and add AI Chat
Left menu → AI ChatOpen in a new tab

Use the dropdown on the left of the input box to select Claude or Codex. The selection is independent per chat and can be switched mid-conversation.

ItemDetails
ProviderClaude (Claude Code CLI) · Codex (OpenAI Codex CLI)
Model selectionSet independently per chat, switchable mid-chat
Multi-turnSession is preserved — context carries over even after restarting the app
ResponseReal-time token-by-token streaming, instantly stoppable with ■ Stop (or Esc)

You can open multiple chats at once, one per panel. Because each chat is an independent session, you can keep working in another chat while one is busy with a long task.

Switch the permission you grant the AI to match the nature of your task, with a single key.

KeyModeWhat the AI can do
Ctrl+1💬 ChatRead only. Analysis, questions, code review. No changes of any kind
Ctrl+2📋 PlanRead + write documents (.md). No code or DB changes — for the design phase
Ctrl+3⚡ Act + ReviewAll tools available, but proceeds only after user approval at every step
Ctrl+4🚀 AutoProceeds autonomously. Ordinary reads/writes are automatic; only dangerous operations require confirmation

Through the MCP bridge, the AI operates the workbench directly with about 40 tools. Tool runs appear inline in the chat thread as tool cards (icon + name + elapsed time), and clicking a card expands its inputs and results.

AreaRepresentative tools
FilesRead/write, directory listing (auto-routed over SFTP for remote projects)
DBRun queries, inspect schema, analyze column statistics
SSH · WinRMRun commands, collect output, auto-detect server environment
Git · ChangeSetstatus/diff/log/commit, checkpoints, revert changes
Run · PreviewRun and observe programs, screenshots, console logs, click/type, visual regression
RAGWorkspace semantic search (workspace_search)
EditorInject code directly (Apply/Reject card in review mode)
OtherIsolated Python execution, screen/VNC capture, ask the user, notifications

Every code block in an AI reply gets a run button.

  • SQL — ▶ runs it in the active DB panel
  • Bash/PowerShell — ▶ runs it in an SSH or terminal panel
  • diff — 🔀 applies it to the file

Run results are automatically attached to the context of the next message.

Computations like data analysis, algorithm verification, and regex testing run directly in Python inside a per-chat isolated venv. The system Python is unaffected, and package installs (pip) go through a confirmation dialog. Deleting a chat also cleans up its venv.

SQL and shell commands invoked by the AI are automatically classified by risk through static analysis.

RiskExamplesBehavior
ReadSELECT, ls, catRuns automatically
WriteINSERT/UPDATE (with WHERE), mv, chmodConfirmation dialog
DestructiveDROP/TRUNCATE, DELETE without WHERE, rm -rfStrong confirmation + impact preview
System paths/etc, C:\WindowsBlocked

Attaching a prod environment label to a DB connection raises that connection’s risk level by one step.

Tool approval gate (⚡ Act + Review mode)

Section titled “Tool approval gate (⚡ Act + Review mode)”

In ⚡ Act + Review mode, the AI shows an approval card before modifying files or running commands.

File edit card

  • Preview of the change as a +/- diff
  • Choose [Apply] [Reject] [Auto-allow for this session]

Command confirmation card

  • Shows the command, working directory, and reason
  • Choose [Run] [Reject] [Auto-allow for this session]

If you reject and enter a reason, it is relayed to the AI so it can propose a different approach. The AI waits indefinitely until you approve.

For models with Claude extended thinking or Codex reasoning enabled, the reply includes a collapsible ”💭 Thinking process” card (collapsed by default). Clicking the card reveals the AI’s full internal reasoning. Claude and Codex use the same UI.

MethodDescription
📎 button or @Attach image, text, or PDF files
Ctrl+LAttach the file open in the editor to the current chat
📷 Screen captureCrosshair-drag to select a region → markup → “Send to AI” (external monitors included)
Web preview ✏️Send the preview panel’s screen to the AI
Screen recordingSend a specific frame to the AI

Images are saved as temporary files and then analyzed by the AI with vision.

Automatic workspace awareness (context chips)

Section titled “Automatic workspace awareness (context chips)”

Every time you send a message, the panel state of the active tab is automatically collected and passed to the AI.

PanelContext passed
EditorVisible code + selection + cursor position
DB result gridThe executed query + result sample + selected cells
DB treeActive connection + schema (tables/views list)
SSH terminalRecent commands + output + selection
SFTPCurrent path + selected files
Folder treeCurrent directory + Git status (branch, modified files)

That’s why phrases like “from this query result,” “the function I’m looking at,” or “on this server” just work.

Panel chips are listed in the “The AI is looking at” bar above the input box.

  • Clicking a chip previews the actual content passed (size, masked fields).
  • Use a chip’s [×] to exclude a specific panel; dimmed (background) panels can be clicked to include them.
  • Size is capped at 4KB per panel and 20KB total, with automatic summarization.
  • Passwords, API keys, tokens, and private keys are automatically masked and never passed to the AI. prod connections are hidden by default.

When multiple chats are open at once, the AI is automatically aware of other chats’ recent work via workspace_recent / workspace_locks. You can share notes between chats with workspace notes (workspace_note). For example, if you write “Production check at 2 AM, no DDL allowed” in one chat, the AI in other chats reads it and proceeds carefully.

Set the project folder at the top of the AI Chat panel.

  • Primary folder — The reference folder. It is the basis for Git checkpoints and remote search.
  • N additional folders — You can designate multiple folders at once, e.g. frontend + backend.
  • Remote SFTP folder — Designate a remote folder as the project folder and file tools auto-route over SFTP.

Large analyses, reports, and refactors can be split off by the AI into background tasks (up to 5 concurrent). The main chat stays usable, and results are pulled in when finished. When you need a quick delegation, the AI hands subtasks to a subagent (ai_subchat) and continues with the results.

  • Chat list — Listed with title, last activity time, message count, and token usage; search finds past chats.
  • Rename — Double-click a chat title to rename it.
  • Copy message — Use each message’s 📋 button to copy its full content.
  • Code block actions — Every code block gets a Copy / ▶ Run button.
  • Each reply header shows the model · elapsed time · token count, and cumulative tokens/cost (USD) are tallied per chat.
  • Chat history is stored locally (userData/chats/). Sensitive information is stored only in masked form.

Before answering, the AI semantically searches the workspace (workspace_search) to reference project code, DB schema, and past chats, marking cited files with “Reference:”. Clicking jumps to the original location so you can verify the basis of the AI’s answer directly. Embeddings are processed locally, so your code never leaves your machine.

For using RAG search directly (shortcut Ctrl+K, source filters, reranking settings, etc.), see the RAG Search doc.

Files, DB, and SSH changes made by the AI are automatically recorded in ChangeSet. From the 📦 ChangeSet chip at the bottom of the chat or via Ctrl+Shift+H, use the “this task” filter to review what the AI did and revert items individually.

The “AI setup needed” message keeps appearing In the wizard, proceed in order: install the CLI → claude login (or codex login). After installing, restart Blyck so the CLI is recognized.

No response, or it stops midway Auto-retry runs up to 3 times. If authentication has expired, re-login and open a new chat. Pressing ■ Stop, then refining and resending the message also helps.

An empty response after stopping and re-requesting This happens when you stop in the middle of a tool run. Open a new chat to continue.

A file changed without an approval card This happens when 🚀 Auto mode (Ctrl+4) is set. Switch to Ctrl+3 (⚡ Act + Review) to get an approval card at every step.

I don’t want to send a specific panel’s info to the AI In the context preview bar, click the [×] on that panel’s chip to exclude it from this message.

Could a password be exposed to the AI? Passwords, API keys, tokens, and private keys are automatically masked and never passed to the AI. prod connections are excluded from context by default.

RAG search returns nothing or something wrong The model may be loading/downloading, or indexing may be in progress. Adjust the source filters, or use .termeditignore to exclude noisy paths. See RAG Search for details.