Use the company DeepSeek V4.1 in your coding agent

1. Quick setup: one command (macOS, Linux, WSL)

Open a terminal and run:

curl -fsSL https://deepseek-dev.7.solutions/setup.sh | bash

1. The script shows a short code and opens your browser.
2. Sign in with Lark, check that the code matches, and click Approve.
3. The script receives a new personal key and sets up the agents it finds on your computer (Claude Code, Codex, Hermes). Your existing settings are not changed.

Then start your agent:

# only some agents
curl -fsSL https://deepseek-dev.7.solutions/setup.sh | bash -s -- --claude --codex
# also make DeepSeek the default model in Hermes
curl -fsSL https://deepseek-dev.7.solutions/setup.sh | bash -s -- --hermes --hermes-default
# remove everything the script added (then revoke the key on the My keys page)
curl -fsSL https://deepseek-dev.7.solutions/setup.sh | bash -s -- --uninstall

2. Manual setup (Windows, or if you prefer)

Get a key Sign in with Lark on the My keys page and create a key. It is shown once; copy it. In the snippets below, replace YOUR_KEY with it. After you create a key, the page shows the same snippets with the key already filled in.

export ANTHROPIC_BASE_URL=https://deepseek-api.7.solutions/anthropic
export ANTHROPIC_AUTH_TOKEN=YOUR_KEY
export ANTHROPIC_MODEL=gce-dsv41/deepseek-v4.1-flash
export ANTHROPIC_DEFAULT_OPUS_MODEL=gce-dsv41/deepseek-v4.1-flash
export ANTHROPIC_DEFAULT_SONNET_MODEL=gce-dsv41/deepseek-v4.1-flash
export ANTHROPIC_DEFAULT_HAIKU_MODEL=gce-dsv41/deepseek-v4.1-flash
export CLAUDE_CODE_SUBAGENT_MODEL=gce-dsv41/deepseek-v4.1-flash
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=262144
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
export API_TIMEOUT_MS=600000
claude

model_provider = "deepseek7"
model = "gce-dsv41/deepseek-v4.1-flash"
model_context_window = 262144

[model_providers.deepseek7]
name = "DeepSeek V4.1 (7 Solutions)"
base_url = "https://deepseek-api.7.solutions/v1"
env_key = "DEEPSEEK7_API_KEY"
wire_api = "responses"

export DEEPSEEK7_API_KEY=YOUR_KEY
codex --profile deepseek7

hermes config set providers.deepseek7.name "DeepSeek V4.1 (7 Solutions)"
hermes config set providers.deepseek7.api https://deepseek-api.7.solutions/v1
hermes config set providers.deepseek7.key_env DEEPSEEK7_API_KEY
hermes config set providers.deepseek7.transport chat_completions
hermes config set providers.deepseek7.default_model gce-dsv41/deepseek-v4.1-flash
hermes config set providers.deepseek7.context_length 262144
hermes config set 'agent.reasoning_overrides.gce-dsv41/deepseek-v4\.1-flash' high
hermes config set DEEPSEEK7_API_KEY YOUR_KEY

curl https://deepseek-api.7.solutions/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" -H "Content-Type: application/json" \
  -d '{"model":"gce-dsv41/deepseek-v4.1-flash","messages":[{"role":"user","content":"Hello"}]}'

Windows PowerShell: set the same variables with $env:NAME = "value" (for example $env:ANTHROPIC_AUTH_TOKEN = "YOUR_KEY"), or permanently with setx NAME value. The Codex profile file goes to %USERPROFILE%\.codex\deepseek7.config.toml.

3. Good to know

• Shared GPU. One server serves everyone: about 4 answers at full speed at once (~35 tokens/s each), and long prompts are read at ~5,000 tokens/s. Please don't run unattended loops with many parallel agents.
• Context. Up to 1M tokens. The snippets set 262,144 because huge contexts are slow and use memory everyone shares; raise it when you really need more.
• Your keys. Up to 5 keys per person (one per computer is a good habit). Each key allows 120 requests per minute and expires after 90 days; run the setup again to get a fresh one. Revoke a lost key right away on My keys.
• Availability. The GPU is a low-cost Spot machine. If Google reclaims it, it restarts on its own; expect errors for about 20–30 minutes, then it works again.
• Privacy. Everything runs in the company Google Cloud project. The gateway does not store prompts or answers.
• Browser chat. For plain chat, use https://deepseek.7.solutions (Lark login, no key needed).

4. Troubleshooting