A simplified Node.js-based agentic IDE for local LLMs via Ollama. Reside understands native JSON tool call formats from Qwen, Granite, DeepSeek, and other models — something most agentic IDE plugins get wrong.
Minimal dependencies. Pure Node.js with optional Puppeteer for web search and content fetching.
- Local & private — Runs entirely on your machine via Ollama. No data leaves your computer.
- Model-native JSON tool calls — Understands Qwen 2.5 Coder (markdown-fenced JSON), Qwen 3.5 (raw JSON with
thinkingfield), Granite 4.1 (raw JSON), and DeepSeek Coder V2 (Unicode XML-like tags with JSON arguments) tool call formats out of the box. Automatically repairs common LLM JSON mistakes like template literals, trailing commas, and single quotes. - 12 built-in tools — Filesystem operations, web search (DuckDuckGo, no API key), URL content extraction, current time, and shell command execution — all with automatic error recovery and self-healing.
- Self-healing — After the LLM writes an app, Reside automatically runs it and injects any runtime errors back into the conversation for the LLM to fix, with up to 3 fix attempts.
- Per-app git repos — Each project directory gets its own independent git repository with auto-commit after every tool execution.
Reside connects to a local LLM running in Ollama, gives it filesystem access through a set of tools, and lets it build projects inside a workdir. Each app/project directory gets its own independent git repository.
You ──> reside ──> Ollama (qwen2.5-coder / qwen3.5 / granite4.1)
│
└──> Workdir/
├── my-app/ (git repo)
│ ├── index.html
│ └── style.css
├── todo-app/ (git repo)
│ └── app.js
└── api-server/ (git repo)
└── server.js
- Node.js v18+ (tested on v24)
- Ollama running locally with at least one model pulled
- Google Chrome (for web search and browser-based URL fetching) —
search_webandfetch_url(browser mode) use Puppeteer with stealth plugin to bypass bot detection. Install Chrome via your package manager or download from google.com/chrome.
# Install a model if you haven't already
ollama pull qwen2.5-coder:7b
ollama pull qwen3.5:latest
ollama pull granite4.1:8b# Clone or create the project
cd reside
# Run a single task
node src/index.js "Create a simple HTML page with a blue button"
# Start an interactive chat session
node src/index.js --chat
# Use a specific model
node src/index.js --chat --model qwen3.5:latestRun a task and exit:
node src/index.js "Create a simple Express.js API server with a /hello endpoint"The agent will:
- Create app directories under
workdir/ - Initialize a git repository in each new app directory
- Process your task through the LLM
- Execute tool calls (write files, run commands, etc.)
- Auto-commit changes to each app's git repo
- Tell you the command to run to start the app (e.g.,
node app.jsin the app directory) - Print a summary and exit
Note: The agent builds your app — it creates source files, installs dependencies, and sets up the project structure. For Node.js apps,
"type": "module"is automatically added topackage.jsonso the LLM usesimport/exportsyntax. The agent does not attempt to run the app; it tells you the exact command to run in your terminal instead.
Start an interactive session where you can have a back-and-forth conversation:
node src/index.js --chatThis opens a REPL prompt where you can type messages, ask follow-up questions, and give sequential instructions. The model remembers the conversation history.
Chat commands:
| Command | Description |
|---|---|
/exit or /quit |
End the session and quit |
/sum or /summary |
Show session summary |
/git or /log |
Show git commit log for all apps |
/apps |
List apps in the workdir |
/clear |
Clear conversation history (keeps system prompt) |
/model |
Show the current model |
/help |
Show available commands |
Example chat session:
$ node src/index.js --chat
📁 Workdir: /home/user/reside/workdir
🤖 Model: qwen2.5-coder:7b
Interactive mode. Type your messages or commands:
💬 > What time is it?
🔧 get_current_time({"format":"full"})
✅ Monday, 28 April 2026 at 10:15:42 British Summer Time (Europe/London, unix: 1777356942)
🤖 It's Monday, 28 April 2026, 10:15:42 BST (Europe/London).
💬 > What's the current price of bitcoin?
🔧 search_web({"query":"current bitcoin price USD"})
✅ Found 5 results
🤖 The current price of Bitcoin is approximately $76,711.52 USD as of the latest data.
💬 > Create a weather CLI app that takes a city name and shows the temperature
🔧 create_directory({"path":"weather-cli"})
✅ Directory created
🔧 write_file({"path":"weather-cli/app.js","content":"import https from 'node:https';\nconst city = process.argv[2] || 'London';\nconst url = `https://wttr.in/${city}?format=%t+%C`;\nhttps.get(url, res => {\n let data = '';\n res.on('data', c => data += c);\n res.on('end', () => console.log(`${city}: ${data.trim()}`));\n});"})
✅ Written 278 bytes to weather-cli/app.js
🔧 test_app({"args":"London"})
✅ London: +12°C Partly cloudy
🔧 finish({"message":"Created weather-cli app that shows temperature for any city"})
✅ Created weather-cli app that shows temperature for any city
💬 > Now modify it to also show the humidity
🔧 read_file({"path":"weather-cli/app.js"})
✅ Read 278 bytes
🔧 edit_file({"file_path":"weather-cli/app.js","new_string":"const url = `https://wttr.in/${city}?format=%t+%C+%h`;\nhttps.get(url, res => {\n let data = '';\n res.on('data', c => data += c);\n res.on('end', () => {\n const [temp, condition, humidity] = data.trim().split(' ');\n console.log(`${city}: ${temp}, ${condition}, Humidity: ${humidity}`);\n });\n});"})
✅ Edited weather-cli/app.js
🔧 test_app({"args":"London"})
✅ London: +12°C, Partly cloudy, Humidity: 65%
🔧 finish({"message":"Updated weather-cli to show temperature, conditions, and humidity"})
✅ Updated weather-cli to show temperature, conditions, and humidity
💬 > /summary
📊 Session Summary:
Iterations: 8
Messages: 28
Workdir: /home/user/reside/workdir
Apps in workdir:
📜 weather-cli
abc1234 Iteration 1: create_directory, write_file, test_app, finish
def5678 Iteration 2: read_file, edit_file, test_app, finish
💬 > /exit
node src/index.js --file task.txt# List all models available in Ollama
node src/index.js --list-models
# List all apps in the workdir (with git status)
node src/index.js --list-apps# Set the default model
node src/index.js --set model qwen3.5:latest
# Set the workdir directory
node src/index.js --set workdir /path/to/projectsConfiguration is stored in ~/.config/reside/config.json. You can also use environment variables:
| Variable | Description |
|---|---|
RESIDE_MODEL |
Default model name |
RESIDE_WORKDIR |
Base workdir directory |
RESIDE_OLLAMA_HOST |
Ollama API host (default: http://localhost:11434) |
The LLM has access to 13 built-in tools. See the full Tools Reference for detailed documentation, parameters, examples, and behavior for each tool.
| Tool | Description |
|---|---|
read_file(path) |
Read the contents of a file |
write_file(path, content) |
Create a NEW file with content (creates directories if needed) |
edit_file(path, file_path, new_string) |
Edit an EXISTING file by replacing text (use write_file for new files) |
list_files(path) |
List files and directories in a path |
search_files(path, regex, file_pattern) |
Search for patterns in files using regex |
create_directory(path) |
Create a directory (and parents if needed); auto-initializes git for top-level app dirs |
execute_command(command, cwd?) |
Run a shell command with timeout and auto-cwd |
delete_file(path) |
Delete a file or directory |
search_npm_packages(query) |
Search the npm registry for packages (use before npm install to verify package names) |
search_web(query) |
Search the web for information (DuckDuckGo, no API key needed) |
fetch_url(url, useBrowser?) |
Fetch a URL and extract its main article content (strips nav, ads, boilerplate) |
get_current_time(format?) |
Get the current system date/time in various formats |
test_app(args?, cwd?) |
Test the application by running it and checking for errors |
finish(message) |
Signal that a task is complete |
To see full raw LLM responses, intermediate logs, and verbose tool output, use the --debug flag:
# Single task with debug output
node src/index.js --debug "Create a simple HTML page"
# Chat mode with debug output
node src/index.js --chat --debugIn debug mode, the agent shows:
- Full raw JSON arguments for each tool call
- Complete tool output (truncated at 500 chars)
- Intermediate logs from search and fetch operations (e.g., "DDG returned N raw results", "Followed N results")
- No compact one-liner summaries — you see the raw output
You can also enable debug mode permanently by setting debugMode: true in ~/.config/reside/config.json.
The workdir is a container directory (default: workdir/). Each app/project you create gets its own subdirectory with an independent git repository:
- Per-app git repos —
create_directory("my-app")auto-initializes git inmy-app/ - Auto-committed — Changes are committed to each app's git repo after each tool execution iteration
- Sandboxed — Path traversal is blocked; tools can't escape the workdir
- Independent history — Each app has its own commit log, branches, and state
- No files in the root — Files must always be written inside an app subdirectory (e.g.,
my-app/app.js), never directly in the workdir root. Thewrite_filetool will reject root-level paths with a clear error message.
# List all apps with git status
node src/index.js --list-apps
# Example output:
# 📜 my-app
# /home/user/reside/workdir/my-app
# a1b2c3d Iteration 1: create_directory, write_file
# e5f6g7h Initial commit
# 📁 todo-app (no git yet)
# /home/user/reside/workdir/todo-appReside is model-agnostic and works with any Ollama model. It has been tested with:
| Model | Tool Call Format | Notes |
|---|---|---|
| qwen2.5-coder:7b | Markdown-fenced JSON (```json {...} ```) |
Well-tested, reliable |
| qwen3.5:latest | Raw JSON (no fences, includes thinking field) |
Well-tested, reliable |
| granite4.1:8b | Raw JSON (no fences, no thinking field) | Well-tested, reliable |
| deepseek-coder-v2:latest | Unicode XML-like tags (<｜tool▁calls▁begin｜>...<｜tool▁calls▁end｜>) with JSON arguments |
Tested and working |
The parser automatically detects and handles all formats. To use a different model:
node src/index.js --model llama3.2:latest "Your task here"src/index.js CLI argument parsing, routes to single-run or chat mode
│
src/config.js Loads config from ~/.config/reside/config.json + env vars
│
src/agent/index.js Agent loop: sends messages to LLM, parses responses,
│ executes tools, feeds results back, detects loops
│
├── src/ollama.js HTTP client for Ollama API
├── src/parser.js Parses JSON tool call format (Qwen 2.5, Qwen 3.5, Granite)
├── src/tools/ Tool execution engine (13 tools, modular)
│ ├── index.js Tool registry and dispatch
│ ├── testApp.js App testing with root + src/ entry point support
│ └── utils/ Shared utilities (workspace, fetchUrl, search, etc.)
└── src/agent/utils/ Agent utilities (briefToolStatus, renderText)
Reside's system prompt includes built-in guidance that tells the LLM to structure generated applications with a controller pattern — separating concerns into dedicated modules rather than writing everything into a single monolithic entry point file.
When the LLM builds an app, it is instructed to follow this structure:
my-app/
├── app.js # Thin entry point — imports, config, server start
├── controllers/ # Route handlers / business logic per domain
│ ├── weatherController.js
│ └── userController.js
├── services/ # Business logic / data access layer
│ ├── weatherService.js
│ └── ...
├── middleware/ # Express middleware (if applicable)
├── models/ # Data models / schemas
├── utils/ # Helper functions
└── config/ # Configuration
Key principles:
app.jsstays minimal (~50 lines max): imports, middleware setup, route mounting, server startup- Controllers handle routing and response formatting for a specific domain
- Services contain the actual business logic (API calls, database queries, file I/O)
- Utils hold pure helper functions (date formatting, validation, string manipulation)
- Each file has a single responsibility; split files that exceed ~200 lines
- For smaller apps (under ~100 lines total), a single file is acceptable
This guidance is embedded in the system prompt at src/config.js and applies to all apps the LLM generates.
Most agentic IDE plugins don't understand native JSON tool call formats from models like Qwen, Granite, and DeepSeek. They expect OpenAI-style function_call structures, but these models output tool calls as JSON objects embedded in the response text — either wrapped in markdown code fences (Qwen 2.5 Coder), as raw JSON with a thinking field (Qwen 3.5), as plain raw JSON (Granite 4.1), or as Unicode XML-like tags with JSON arguments (DeepSeek Coder V2).
Reside was built specifically to handle these formats correctly, with zero dependencies and maximum efficiency.