deepagent-ltd/deepagent-code

DeepAgent Code – An AI‑powered coding assistant

What it is – DeepAgent Code is a self‑hosting workspace that lets a large‑language‑model act as a coding partner. It keeps a durable, version‑controlled memory of the project, can plan and execute multi‑step tasks, and lets a human steer the work in real time.

How it works – The system stitches together four graphs:

1. Code graph – files, symbols, call‑chains, diagnostics (via LSP).
2. Knowledge graph – reusable strategies, facts, and domain‑specific skills.
3. Project memory – decisions, constraints, and learned conventions that persist across sessions.
4. Document graph – plans, work‑logs, evaluations, and evidence. When a task is run, a Session V2 runner pulls relevant pieces from these graphs, admits only vetted snippets to the LLM prompt, and records everything it used as evidence. The LLM then generates code or suggestions, which are committed through a Git‑style workflow that can involve multiple specialist sub‑agents and an optional expert‑panel review.

Collaboration modes

Mode	You give	DeepAgent does
Auto	A single request	Defines the goal, designs a plan, and runs it end‑to‑end.
Loop	A high‑level goal	Writes a goal+plan.md, then iterates through plan → execute → verify → iterate ticks, letting you edit the plan on the fly.
Design	Your pre‑written goal+plan.md	Executes exactly that plan without redefining objectives.

Control & safety – You can pause, resume, roll back, or take over any workflow. Guidance can be injected live (while a tool is running) or at the next tick, and all actions are logged in an immutable audit trail.

Key technical pieces

• AI IDE microservice – Symbol‑and‑intent queries over 38 language‑server back‑ends, with unsaved buffers included.
• Domain packs – Plug‑in knowledge for languages, frameworks, hardware, business rules, etc., that activate automatically based on the problem profile.
• Specialist agents & Expert Panel – Isolated workers can handle sub‑tasks; a panel can review high‑risk decisions anonymously before an arbiter merges the result.
• Provider‑agnostic model support – 75+ model providers via the AI‑SDK, plus any OpenAI/Anthropic‑compatible endpoint.
• Durable memory – Typed, versioned documents (knowledge, strategy, memory, design, worklog…) with provenance, confidence scores, and lineage.

Getting started

# Install (macOS / Linux) – runs a script that sets up the desktop app and CLI
curl -fsSL https://deepagent.ltd/install | bash

# Run the CLI
deepagent-code            # or the alias `deepagent`

Add a model provider through the desktop Settings → Providers UI, or via the CLI:

deepagent auth login      # opens a flow for OpenAI, Anthropic, etc.
deepagent auth list       # shows configured providers

You can also edit ~/.deepagent/code/config.jsonc to add custom OpenAI‑compatible endpoints.

Example usage

deepagent-code run "add rate limiting to /api/users endpoint"

The agent will locate the endpoint via LSP, reuse any existing middleware patterns from project memory, activate the backend‑API domain pack, implement the rate‑limiter, run tests, and store the new pattern as a reusable memory entry.

Build from source (requires Bun 1.3.14)

git clone https://github.com/deepagent-ltd/deepagent-code.git
cd deepagent-code
bun install
bun run dev:desktop   # launch the desktop UI
bun run dev            # launch the terminal UI

License – AGPL‑3.0‑or‑later (network‑service deployments must share source). The core is derived from the MIT‑licensed opencode project.

---

DeepAgent Code is positioned as a production‑grade AI coding agent that remembers past work, plans complex tasks, lets humans intervene safely, and keeps a transparent, versioned knowledge base.