Skip to content
DeepCode

Getting started

Fifteen minutes from nothing to a working session. Three steps: install, connect one model, run.

1 · Install

Section titled “1 · Install”
Terminal window
uv tool install --python 3.12 deepcode-hku
deepcode init

pip install deepcode-hku and pipx work too, in a Python 3.12+ environment. deepcode init creates ~/.deepcode/ — configuration lives there, credentials in private storage beside it, never inside conversation history.

The Desktop app installs separately — grab a release bundle from GitHub Releases. Everything in these guides about sessions, skills, and permissions applies to it too.

2 · Connect a model

Section titled “2 · Connect a model”

One connection is enough to start. --api-key opens a hidden prompt — the key never appears in your shell history:

Terminal window
deepcode provider set my-openrouter --template openrouter --api-key
deepcode provider models my-openrouter --refresh
deepcode provider test my-openrouter --model deepseek/deepseek-v4-pro

Read that as: create a connection → see what it can serve → prove one model answers. Templates exist for OpenAI, Anthropic, OpenRouter, DeepSeek, Gemini, Ollama, vLLM and more; deepcode provider list shows what’s configured. (Source of truth: cli/provider_cli.py.)

3 · Run

Section titled “3 · Run”
Terminal window
cd <your-project>
deepcode

The first time you point DeepCode at a folder, it asks before touching anything:

DeepCode can read files and run tools in this workspace:
  /Users/you/your-project
Trust this folder? Type yes to continue:

Trust is per-folder and remembered. Then the banner:

 ██████  ·──○
 ██       ○──·
 ██████  ·──○
 DeepCode · open agentic coding
 deepseek/deepseek-v4-pro · ~/your-project
 session 3f2a91c0 · access default (ask) · effort auto
 /help for commands · esc interrupts · ctrl+o transcript detail

That third line is your session’s identity: the model it will use, the session id you can resume later, and how much the agent may do without asking.

4 · Read your first turn

Section titled “4 · Read your first turn”

Type a task in plain language:

› find where retries are configured and raise the limit to 5

The reply is a rhythm of cards you’ll see constantly, so learn it once:

● Search retry
  ⎿ ✓ 3 files matched
● Read src/client/http.py
  ⎿ ✓ 1: import httpx
● Edit src/client/http.py
  ⎿ ✓ http.py


Raised MAX_RETRIES from 3 to 5 in src/client/http.py.
· 12s · 9.4k in · 210 out
  • ● — a tool starts: what the agent is doing (Read, Search, Run, Edit…) and on what.
  • ⎿ ✓ / ⎿ ✗ — that tool settled, with the first line of its result.
  • Plain text — the agent talking to you.
  • The dim · 12s · 9.4k in · 210 out footer — the turn is finished: wall time, prompt tokens in, completion tokens out. Real numbers from the provider, not estimates.

When the agent wants to do something sensitive under the default access preset, it stops and asks:

◆ approval needed bash
  pip install requests
  ⎿ reply y once · a session · n deny

Reply y (this once), a (for the whole session), or n. Nothing sensitive runs without an answer.

Where to next

Section titled “Where to next”