DeepAgent Code (DAC)

What it is

DeepAgent Code (DAC) is a desktop‑plus‑terminal workspace that lets you use a large‑language‑model (LLM) as a coding assistant for tasks that require many back‑and‑forth steps. It combines:

• a code editor, terminal, and LSP‑based navigation,
• a persistent session that remembers project state, goals and plans across model turns,
• a project memory that you can inspect and edit instead of hidden prompt history,
• sub‑agents that can run independent jobs in isolated work‑trees, and
• a managed browser so the agent can read or interact with web pages without leaving the workspace.

All of these pieces share a single execution core, and you can choose any compatible LLM provider (OpenAI, Anthropic, DeepAgent API, etc.) and configure tools (called Skills and MCP connectors) that the model can invoke.

Why you might care

Pain point	How DAC helps
Long, multi‑step coding tasks	Sessions keep the conversation, goals, and plan alive; you can steer the model between provider turns and queue follow‑up tasks.
Repeating project context	Project memory stores facts, decisions, and tool results with provenance, so the model only gets the pieces it needs, not the whole repo each time.
Splitting work among helpers	Sub‑agents run isolated jobs, keep durable run records, and report back, letting you delegate research or implementation without losing control.
Needing web information	A desktop‑managed browser lets the agent inspect pages, download/upload files, and interact with the clipboard under explicit user authorisation.
Turning chat into code	In supported instant‑messaging (IM) platforms you can invoke /dac … to start a session, watch execution in the DAC UI, and get the final answer back in the chat thread.
Tooling flexibility	GUI and TUI share the same core; you can plug in different LLM providers, add custom Skills, domain packs, or MCP connectors.

Core concepts

Concept	Description
Session	Persistent execution context. Inputs are saved before they are sent to the model; you can edit goals, add plans, and steer the session at safe boundaries.
Project memory	Structured store of code references, decisions, tool results, and versioned facts. Editable and inspectable, it feeds the model on demand while respecting token budgets.
Sub‑agents	Independent agents that run in their own work‑trees. Their runs are recorded with ownership, status, and results, enabling collaborative workflows.
Event workflow layer	Configurable events that build dependency‑aware task graphs, automatically tracking what triggered each step and what still needs attention.
Skills / MCP	Reusable procedures (Skills) and external tool connectors (MCP) that the LLM can call, all governed by permission checks.
IM integration	/dac commands in supported instant‑messaging apps launch a local session, stream execution in the DAC UI, and return the final answer to the chat.

Getting started (as described in the README)

1. Download a pre‑built desktop package from the GitHub releases page (current line 2.0.4).
2. Install the model provider in Settings → Providers (e.g., DeepAgent API or your own API key).
3. Open a project and start a conversation – the UI shows an editable plan and a live stream of model output.
4. Optional – build from source using Bun 1.4.2:

   git clone --branch 2.0.4 https://github.com/deepagent-ltd/deepagent-code.git
   cd deepagent-code
   bun install
   bun run dev:desktop   # desktop UI
   bun run dev           # terminal UI
   bun run dev:web       # web UI (needs backend)
   

5. Use IM (if you have the paid IM add‑on) by typing /dac <task> in a chat; DAC runs locally and posts the answer back.

License & community

• License: AGPL‑3.0‑or‑later (copyleft).
• Contributions: welcomed via the usual pull‑request flow; see CONTRIBUTING.md.
• Security & privacy: policies are provided in the repo.
• Built on the open‑source opencode project and other components, with proper attribution.

---

DeepAgent Code is a genuine, actively maintained software project that provides an AI‑augmented coding workspace for complex, multi‑turn development tasks.