CCE — Code Chat Extension
A minimal, transparent VSCode chat interface for LLMs. You decide what gets sent to the model. No hidden prompts, no magic context — just a clean channel between you and the model, with tools the model can opt into.
Philosophy
- Full control. Which files, selections, and system prompts reach the model is always your choice. Nothing is injected behind your back.
- Tools are explicit. Every tool the model can use — read files, search code, run commands — is listed, toggleable, and inspectable. The model must explicitly request each one.
- Minimal by design. No framework bloat, no bundler, no abstractions over abstractions. Vanilla JavaScript talking to LLM APIs. What you see is what you get.
Features
- Multi-model — OpenAI, Anthropic, and any OpenAI-compatible endpoint (Ollama, vLLM, LM Studio)
- Streaming responses — tokens appear as they're generated
- Markdown rendering — code blocks, tables, lists via
marked
- Thinking display — reasoning content in collapsible blocks (o1, o3, DeepSeek-R1, Qwen)
- Tool calling — model can read/write files, search code, list directories, run commands, get diagnostics
- MCP servers — Model Context Protocol support for external tools (web search, APIs)
- Agent tool — spawn parallel sub-agents for independent tasks
- Sessions — persistent conversation history with full session management
- Config panel — tabbed settings UI for models, context, tools, and MCP servers
- API keys in OS keychain — stored via VSCode
SecretStorage (Windows Credential Manager, macOS Keychain, Linux libsecret)
Quick Start
Prerequisites
Install locally
git clone https://github.com/5im-0n/cce
cd cce
npm install
npx vsce package
This creates cce-0.0.1.vsix. Install it in VSCode:
code --install-extension cce-0.0.1.vsix
Or via VSCode UI: Extensions → ... → Install from VSIX...
Run in development
Open the project in VSCode and press F5. This launches an Extension Development Host with CCE loaded.
Configuration
Open CCE from the chat icon in the activity bar on the left. Click the gear icon to open settings.
Models
Add models via Settings → Models. Each model needs:
| Field |
Description |
| Display Name |
Shown in the model dropdown |
| Provider |
OpenAI / OpenAI-compatible (Anthropic coming soon) |
| Model ID |
API model identifier (e.g. gpt-4o, claude-sonnet-4) |
| API Endpoint |
Base URL (e.g. https://api.openai.com/v1) |
| API Key |
Optional — leave empty for local models |
The first model added becomes the default.
Context
Control exactly what the model sees:
- System Prompt — custom instructions sent at the start of every conversation
- Auto-inject — optional toggles to include date/time, OS, regional settings, and
AGENTS.md
- AGENTS.md — optionally include a project instructions file (default:
AGENTS.md in workspace root)
Everything is opt-in. The model only sees what you explicitly enable.
Enable only the tools you want the model to have access to:
| Tool |
Description |
| Read File |
Read workspace files with line numbers |
| Write File |
Create, overwrite, insert, replace, or delete lines |
| Get Selection |
Current editor selection and cursor position |
| Search Code |
Literal, case-insensitive workspace search |
| List Files |
Directory listing |
| Get Diagnostics |
VSCode errors, warnings, hints |
| Delete File |
Remove files or directories |
| Agent |
Spawn parallel sub-agents |
| Run Command |
Execute shell commands |
Every tool call is gated by an approval policy before it executes:
- Read-only tools (Read File, Get Selection, Search Code, List Files, Get Diagnostics) run automatically.
- Mutating tools (Write File, Edit File, Delete File) and arbitrary tools (Run Command, Agent, all MCP tools) require your approval.
- When approval is needed, an approval card appears inline in the chat showing the tool name, its arguments, and its risk level, with these choices:
- Allow — run this one call
- Allow for this session — run this tool without asking for the rest of the session (not offered for MCP tools)
- Always allow this tool — persist "auto" for this tool, revocable in Settings → Tools (for MCP tools this becomes "Always allow this server")
- Deny — refuse; the model is told the call was denied and must adapt
- MCP approval is server-level. An Auto / Ask / Deny dropdown on each server card in Settings → MCP applies to every tool that server exposes. Per-tool and per-session overrides do not apply to MCP tools; a stale per-tool entry is ignored.
- Approving the Agent tool trusts everything its sub-agents do (they run without further prompts).
- Pending requests are auto-denied after 5 minutes, and denied immediately if you cancel the message, switch sessions, or close the chat.
- Per-tool overrides (Auto / Ask / Deny) are available in Settings → Tools, next to each tool toggle.
MCP Servers
Add Model Context Protocol servers via Settings → MCP. Each server needs a URL and optional custom headers. Toggle servers on/off per-session.
Keyboard Shortcuts
| Key |
Action |
| Enter |
Send message |
| Shift+Enter |
Newline in input |
Security
- API keys stored in OS keychain via
vscode.SecretStorage — never in plaintext files
- Webview never sees API keys
run_command and delete_files are available but disabled by default — enable them explicitly in Settings → Tools- Mutating, shell, and MCP tools require inline user approval before executing (see Tool Approval above)
- Sub-agents are blocked from spawning more sub-agents (no recursive agent explosion)
License
MIT
| |
