Skip to content
BranchesTags

Repository files navigation

PocketAI

PocketAI for VS Code

An agentic AI coding assistant powered by local language models.
Works with any OpenAI-compatible API — Ollama, llama.cpp, LM Studio, vLLM, and more.

pocketaihub.com · Getting Started · Features · Configuration · License

Getting Started

  1. Install the PocketAI 1.0.20 VSIX with VS Code's Extensions: Install from VSIX... command (or build from source)
  2. Start a local model server — e.g. ollama serve
  3. Open PocketAI with Cmd+L (or Ctrl+L on Windows/Linux)
  4. Chat — ask it to read, edit, refactor, debug, or explain your code

PocketAI connects to http://127.0.0.1:39457 by default. Add additional endpoints in settings.

Using Codex CLI

PocketAI now includes a built-in Connect to Codex section in the settings sidebar.

  1. Open the PocketAI sidebar
  2. In the settings view, click Connect to Codex
  3. If you are not signed in yet, PocketAI opens a VS Code terminal and runs codex login for you
  4. When sign-in completes, PocketAI refreshes the Codex endpoint automatically

What this setup does for you:

  • Adds a Codex Bridge endpoint at http://127.0.0.1:39458
  • Starts the local compatibility bridge automatically
  • Switches PocketAI to that endpoint for chat
  • Keeps Codex chat-first without changing tool behavior for your other endpoints

Manual fallback:

codex login
npm run codex-bridge

Notes:

  • apiKey is ignored by the bridge. Authentication comes from your local codex login session.
  • PocketAI remains the executable tool system. The bridge disables Codex-native tools and translates PocketAI tool requests back into PocketAI approvals, inline diffs, and tool execution.
  • PocketAI may still send temperature, top_p, and max_tokens, but Codex app-server does not expose Chat Completions-style tuning controls, so those values are not forwarded 1:1.
  • If you want Codex to use a different workspace root, you can still launch the bridge manually with CODEX_BRIDGE_CWD=/path/to/project npm run codex-bridge.

Using Claude CLI

PocketAI also includes a built-in Connect to Claude section in the settings sidebar.

  1. Open the PocketAI sidebar
  2. In the settings view, click Connect to Claude
  3. If you are not signed in yet, PocketAI opens a VS Code terminal and runs claude auth login for you
  4. When sign-in completes, PocketAI refreshes the Claude endpoint automatically

What this setup does for you:

  • Adds a Claude Bridge endpoint at http://127.0.0.1:39460
  • Starts the local compatibility bridge automatically
  • Switches PocketAI to that endpoint for chat
  • Keeps PocketAI in charge of tool calling while Claude provides the assistant responses

Manual fallback:

claude auth login
npm run claude-bridge

Notes:

  • apiKey is ignored by the bridge. Authentication comes from your local claude auth login session.
  • PocketAI remains the executable tool system. The bridge disables Claude-native tools and translates PocketAI tool requests back into PocketAI approvals, inline diffs, and tool execution.
  • The bridge runs Claude in print mode with Claude-native tools disabled and slash commands turned off, so the experience stays consistent with PocketAI's own harness.
  • If you want Claude to use a different workspace root, you can still launch the bridge manually with CLAUDE_BRIDGE_CWD=/path/to/project npm run claude-bridge.

Using DeepSeek API

PocketAI includes a built-in Connect to DeepSeek API section in the settings sidebar.

  1. Open the PocketAI sidebar
  2. In the settings view, paste your DeepSeek API key and click Save Key
  3. Click Connect to DeepSeek
  4. Pick deepseek-v4-pro or deepseek-v4-flash from the model selector

What this setup does for you:

  • Adds a DeepSeek API Bridge endpoint at http://127.0.0.1:39464
  • Starts a local OpenAI-compatible proxy to https://api.deepseek.com
  • Stores the API key in VS Code secret storage
  • Keeps PocketAI in charge of tool execution while DeepSeek provides model responses

Manual fallback:

export DEEPSEEK_API_KEY="sk-..."
npm run deepseek-bridge

Notes:

  • The bridge exposes DeepSeek V4 model IDs: deepseek-v4-pro and deepseek-v4-flash.
  • Reasoning controls are available as high and max; when selected, the bridge enables DeepSeek thinking mode for the request.
  • The bridge uses port 39464 to avoid the existing Hermes bridge on 39463.

Features

Agentic Tool Loop

PocketAI doesn't just generate text — it acts. The model can chain together file reads, edits, terminal commands, searches, and git operations autonomously, looping until the task is done.

Three Interaction Modes

Mode Behavior
Ask Presents each change for approval before applying
Auto Non-destructive tools run automatically; destructive ones still require approval
Plan Describes what it would do without making any changes

Switch modes with /mode ask, /mode auto, or /mode plan.

Inline Diffs

Pending edits appear directly in your editor with highlighted removed lines and ghost-text previews of replacements. Accept or reject changes via CodeLens buttons without leaving the file.

Integrated Terminal

Shell commands execute in VS Code's integrated terminal so you can see exactly what's running. Output is captured and fed back to the model for the next step.

Context-Aware

PocketAI automatically gathers context with each message:

  • Workspace file tree and active file contents
  • Open editor tabs
  • Git branch and status
  • VS Code diagnostics (errors and warnings)

All configurable — disable any context source in settings.

More

  • MCP Servers — Connect external tool servers via Model Context Protocol for additional capabilities
  • Multi-Session — Create, fork, search, and switch between independent chat sessions
  • Checkpoints & Undo — Every file edit creates a checkpoint; rewind code and conversation at any time
  • Project Instructions — Drop a .pocketai.md in your repo root to customize system behavior per-project
  • Hooks — Run shell commands on lifecycle events (postEdit, preToolUse, postCreate, etc.)
  • Permission Rules — Allow or deny tool calls with glob patterns (e.g. deny run_command(rm *))
  • Session Compaction — Auto-summarizes long conversations to stay within your model's context window
  • @ Mentions — Reference files with @filename autocomplete in the chat input
  • Web Search — The model can search the web when it needs external information
  • Image Support — Paste or attach images to your prompts (for multimodal models)

Configuration

All settings live under pocketai.* in VS Code settings.

Endpoints & Models

Setting Default Description
pocketai.endpoints Local Ollama List of OpenAI-compatible API endpoints with per-endpoint model, token limit, and system prompt
pocketai.useStructuredTools true Use OpenAI function calling instead of text-based tool parsing (disable if your model does not support it)
pocketai.contextWindowSize 8192 Your model's context window size — used by the token meter and auto-compaction

Behavior

Setting Default Description
pocketai.useIntegratedTerminal true Run commands in VS Code's terminal instead of a background process
pocketai.includeGitContext true Include git branch and status in context
pocketai.includeDiagnostics true Include VS Code errors/warnings in context
pocketai.projectInstructionsFile .pocketai.md Filename for per-project system prompt instructions
pocketai.maxContinuations 3 Max auto-continuations when the model hits its token limit mid-response

MCP Servers

"pocketai.mcpServers": [
  {
    "name": "filesystem",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/dir"]
  }
]

Permission Rules

"pocketai.permissions": {
  "allow": ["read_file(**)", "grep(**)", "glob(**)"],
  "deny": ["read_file(*.env)", "run_command(rm *)"]
}

Slash Commands

Command Description
/clear Clear the conversation
/model <name> Switch model
/mode <ask|auto|plan> Switch interaction mode
/compact Summarize conversation to free context
/fork Fork the current session into a new tab
/branch [name] Show, create, or switch git branches
/tokens Show cumulative token usage for the session
/endpoint [name] Switch between configured endpoints
/sessions List all sessions

Keyboard Shortcuts

Shortcut Action
Cmd+L / Ctrl+L Focus chat input
Cmd+Shift+L / Ctrl+Shift+L Send current selection to chat

Building from Source

git clone https://github.com/PocketAIHub/pocketai-vscode.git
cd pocketai-vscode
npm ci
npm run typecheck
npm test
npm run package:vsix

Then press F5 in VS Code to launch the extension in a development host. Run npm run test:extension for the isolated Extension Host suite, or npm run release:check for the complete local release gate.

License

MIT

About

No description, website, or topics provided.

Resources

Readme
MIT license
Activity
Custom properties

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages