Coding Agent

The coding agent works on a folder. It receives a task, reads and edits the files, runs your test command, then repeats until the tests pass or the iteration limit is reached. All file access is confined to the workspace you provide, and it runs on any chat provider.

Tools

The agent calls a small toolset on your workspace:

Tool	Purpose
read_file	Read a file from the workspace.
write_file	Create or replace a file.
edit_file	Replace an exact, unique block of text inside a file.
list_files	List the workspace files.
search	Search the workspace for a regular expression.
bash	Run a command, for example the test suite.

Settings

• apiKey: the provider key.
• provider: any chat provider, such as openai (default), anthropic, gemini, mistral, cohere or ollama.
• model: the model id, for example claude-sonnet-5 or gpt-5.5.
• workspace: the folder the agent is allowed to work in.
• maxIterations: cap on model turns, default 20.
• allowBash: set to false to disable command execution, default true.
• bashTimeout: milliseconds allowed per command, default 120000.
• options: chatbot options such as baseUrl, timeout or retries.
• log: print each tool call when true.
• onAction: a callback ({ iteration, tool, args, thought }) called before each tool runs.

Example

const { CodingAgent } = require('intellinode');

const agent = new CodingAgent({
  apiKey: ANTHROPIC_API_KEY,
  provider: 'anthropic',
  model: 'claude-sonnet-5',
  workspace: './my_repo',
});

const result = await agent.run('Fix the failing tests in calc.js', { testCommand: 'npm test' });

console.log(result.success, result.summary);

run() resolves with { success, summary, iterations, testOutput }. Without a testCommand the agent finishes when the model reports the task done; with one, finish is accepted only when the command exits with code 0, otherwise the test output goes back to the model and the loop continues.

Watching the agent

const agent = new CodingAgent({
  apiKey: OPENAI_API_KEY,
  provider: 'openai',
  workspace: './my_repo',
  maxIterations: 12,
  onAction: ({ iteration, tool, args }) => console.log(`#${iteration} ${tool}`, tool === 'write_file' ? args.path : args),
});

The workspace toolkit

The same tools are available on their own through WorkspaceToolkit, for example to build your own agent loop:

const { WorkspaceToolkit } = require('intellinode');

const tools = new WorkspaceToolkit('./my_repo', { allowBash: true, bashTimeout: 60000 });
console.log(tools.listFiles());
console.log(tools.search('TODO'));
console.log(tools.runBash('npm test'));   // '[exit code 0]\n...'

Every path is resolved inside the workspace; a path that escapes it (including through a symlink) is rejected.

Notes

• The agent is Node only; it is not part of the browser bundle.
• Small local models may not follow the tool protocol reliably; use a capable model for real repositories.