UnieAI Code

A locally-run coding agent for your own models — terminal, editor, or embedded.

UnieAI Code is a coding agent that navigates real repositories — it explores your code, runs commands, edits files, and verifies its own changes — running against your own UnieAI Studio model gateway (Qwen, MiniMax, GLM, and other open models). Nothing leaves your environment except the model calls you already make.

It comes in three forms, all sharing one engine:

| Surface | What it is | |---|---| | CLI (unieai) | A fast Rust terminal agent — interactive TUI or scripted unieai exec for CI. | | Engine (@unieai/code-agent-runtime) | The coding runtime as an npm library, to embed in your own tools. | | VS Code extension | An in-editor chat panel powered by the same engine. |

Internally the engine runs on @unieai/agent-core — our shared agent loop, also used by UnieAI Studio.

Why it's good

The agent harness matters as much as the model. On SWE-bench Lite (300 real GitHub issues, single attempt, official test judging) with a 35B open model, iterating on the harness took the resolved rate from 27.7% → 46.0% — while staying ~2× cheaper per task than a stock codex-style harness, and with no regression on HumanEval (99%). Full data and methodology in benchmarks/RESULTS.md.

What drives it: a completion contract that refuses "done" until the work is actually there and verified, tools that report a syntax check the moment you edit, fuzzy-matching edits that survive whitespace drift, and open-model resilience (empty-response resampling, <think> stripping, stall detection).

What you get

| Feature | Description | |---------|-------------| | Interactive TUI | Terminal UI with markdown, thinking blocks, streaming tool output | | Non-interactive mode | unieai exec / unieai review for CI/CD and automation | | Your models | UnieAI Studio gateway — cloud or on-prem, open models | | VS Code panel | In-editor chat that reads your workspace and edits files | | Sandboxing | Shell commands run in a sandbox; risky actions ask first | | Verification built in | Edits self-check; the agent proves its work before finishing | | MCP integration | Run and manage external MCP servers as tools | | Skills & plugins | Extensible skill definitions and a plugin system | | Sessions | Resume, fork, archive, and delete past sessions |

Quick Start

Install the CLI

macOS / Linux:

curl -fsSL https://www.unieai.com/code/install.sh | sh

Installs the unieai binary to ~/.local/bin. (Windows and pinned-version install options: see docs/install.md.)

Run

unieai                          # interactive session
unieai "fix all type errors"    # start with a prompt

Sign in

unieai login --studio-url https://studio.demo.unieai.com

Opens a device-code flow in your browser; once confirmed, your model list is saved to ~/.unieai/unieai.json. For on-prem deployments, point --studio-url at your Studio endpoint — the inference gateway is auto-derived (studio. → api.).

Use in VS Code

Install the UnieAI Code extension from the marketplace, sign in when prompted, and open the panel from the activity bar.

Embed the engine

npm install @unieai/code-agent-runtime

See agent-runtime/README.md for the API.

Commands

unieai [OPTIONS] [PROMPT]        # interactive TUI

| Command | Description | |---------|-------------| | unieai exec PROMPT | Run the agent non-interactively | | unieai review --diff PATH | Non-interactive code review | | unieai resume / fork | Resume or fork a previous session | | unieai login / logout | Manage UnieAI Studio credentials | | unieai mcp | Manage external MCP servers | | unieai sandbox | Run a command in the sandbox | | unieai doctor | Diagnose install, config, auth, and runtime health | | unieai update | Update to the latest version |

Configuration

CLI config lives in ~/.unieai/config.toml (approval policy, sandbox, models, MCP servers, skills). See the Config Reference; run unieai doctor to diagnose issues.

Repository Structure

├── codex-rs/                     # Rust CLI (TUI, exec, sandbox, app-server)
├── agent-runtime/                # @unieai/code-agent-runtime — the coding engine (npm)
│   ├── src/engine.mjs            #   one turn on the agent-core loop + coding layer
│   └── src/tools.mjs             #   bash / read / write / edit (sandboxed, verifying)
├── third_party/
│   └── unieai-agent-core/        # @unieai/agent-core — shared agent loop (submodule)
├── sdks/vscode/                  # VS Code extension
├── benchmarks/                   # HumanEval + SWE-bench harness and results
├── docs/                         # Documentation
└── sdk/                          # Python and TypeScript SDKs

Contributing

See docs/contributing.md for setup and build instructions. The CLI is a Rust project; the engine is Node ≥ 20. Clone with submodules:

git clone --recursive https://github.com/UnieAI/unieai-code.git

License

Apache-2.0 — see LICENSE.