Inside 100X Code: How Our Agentic CLI Actually Works

Not Autocomplete. An Agent.

Most AI coding tools complete the line you are typing. 100X Code does the task. You describe an outcome, something like "add rate limiting to the API and cover it with tests," and the CLI plans the steps, edits the files, runs the test suite, reads the failures, and iterates until it is done. It works the way a capable engineer works, from a terminal you already live in.

The Agentic Loop

Under the hood, every request runs a plan, act, observe loop:

• Plan: break the goal into concrete steps, scoped to the files and tools involved.
• Act: edit code, run a command, search the repo, or call an external tool.
• Observe: read the result, whether a test output, a compiler error, or a command exit code, and decide the next step.

The loop is what separates a suggestion from a finished change. It is also the heart of the harness layer, the software that turns raw open weights into a system that does real work.

Parallel Sub-Agents and the Orchestrator

Big tasks fan out. 100X Code can spawn focused sub-agents, one writing tests, another updating docs, a third refactoring a module, while an orchestrator coordinates them, merges their work, and keeps them from stepping on each other. You get the throughput of a team without leaving the prompt.

Tools Over MCP

An agent is only as useful as what it can reach. 100X Code connects to your systems through the Model Context Protocol: your ticket tracker, your database, your internal wiki, your cloud console. When a task genuinely needs the open web, it can search, but only when you allow it, and never with your private context attached. The connections are yours to grant and revoke.

Many Sessions, One Memory

Real work is not linear. You can run several sessions in parallel, a long refactor in one and a quick bug fix in another, and Chitta keeps a shared, permission-aware memory across them, so a decision made in one session is not forgotten in the next.

It Runs on Your Model

Every one of these capabilities points at an engine you control, Pro on your own hardware or the managed, air-gapped Flash. The agent is powerful precisely because it is not a black box in someone else's cloud. When you are ready, the quickstart gets you running.