deepseek-agent 0.6.2

Developer-focused DeepSeek command line tools

DeepSeek CLI

This CLI is open source at https://github.com/PDFSage/deepseek_cli – collaborators and maintainers are welcome! Submit ideas, issues, or pull requests to help the project grow.

https://pypi.org/project/deepseek-agent/

Developer-focused command line tools for working with DeepSeek models. The CLI packages both an interactive chat shell and an agentic coding assistant with repository-aware tooling, plus configuration helpers and transcript logging.

Features

• Unified agent shell orchestrates tool-aware coding sessions with the DeepSeek API, optional read-only mode, transcripts, and workspace controls. It launches by default when you run deepseek.
• Inline shortcuts (@chat, @complete, @embed, @models) surface chat, completion, embedding, and model-list primitives without leaving the agent shell.
• Auto-detects likely project test commands and reminds the agent to run them, keeping changes validated against the repo's real workflows.
• Config mode (deepseek config) manages stored defaults while respecting environment variable overrides.
• Interactive mode now launches a colourful Rich-powered shell that surfaces / and @ command shortcuts, streams progress spinners for non-shell tool calls, and lets you update the stored API key without leaving the session.
• Built-in Tavily web search support supplements the agent with live research, using a bundled developer key that you can override per session or in config.
• Auto-router smart search mode (deepseek autorouter) heuristically pulls Tavily context, then streams DeepSeek Reasoner's thinking/status via SSE.
• Claude Code-inspired workflows (--workflow feature-dev, --workflow pr-review) and personas (--persona code-architect, --persona code-explorer, --persona code-reviewer) can be layered onto any run or toggled live via @workflow/@persona, giving the agent guided phase plans and specialist voices without leaving the CLI.
• Standalone Tavily utilities (deepseek tavily search …, deepseek tavily extract …) plus new tavily_extract/todo_write tools let the agent research up-to-date docs and keep structured TODO logs inside the workspace.
• Ships as a Python package with an executable entry point and Homebrew formula for distribution flexibility.

Requirements

• Python 3.9 or newer.
• A DeepSeek API key exported as DEEPSEEK_API_KEY or stored via deepseek config.
• pip 23+ is recommended. Create a virtual environment for isolated installs.

Installation

From PyPI (recommended once released)

python -m pip install --upgrade pip
python -m pip install deepseek-agent

To update later, run python -m pip install --upgrade deepseek-agent.

From GitHub

Install the latest commit directly from GitHub:

python -m pip install "git+https://github.com/PDFSage/deepseek_cli.git@main"

Specify a tag (for example v0.2.0) to pin a release:

python -m pip install "git+https://github.com/PDFSage/deepseek_cli.git@v0.2.0"

From a local clone

git clone https://github.com/PDFSage/deepseek_cli.git
cd deepseek_cli
python -m venv .venv
source .venv/bin/activate  # On Windows use: .venv\\Scripts\\activate
python -m pip install --upgrade pip
python -m pip install -e .  # or `python -m pip install .` for a standard install

The editable install (-e) keeps the CLI synced with local source changes while developing.

Configuration

The CLI resolves settings in the following order:

1. Command line flags (--api-key, --base-url, --model, etc.).
2. Environment variables: DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL, DEEPSEEK_MODEL, DEEPSEEK_SYSTEM_PROMPT, DEEPSEEK_CHAT_MODEL, DEEPSEEK_COMPLETION_MODEL, DEEPSEEK_EMBEDDING_MODEL, DEEPSEEK_CHAT_STREAM_STYLE, TAVILY_API_KEY.
3. Stored configuration file at ~/.config/deepseek-cli/config.json.

Values from a local .env file are automatically loaded (via python-dotenv) before reading environment variables, so you can keep API keys there if you prefer not to export them globally.

Helpful commands:

deepseek config init        # Guided prompt to store your API key
deepseek config show        # Display the current configuration (API key redacted)
deepseek config show --raw  # Show the API key in plain text
deepseek config set model deepseek-reasoner  # Update an individual field
deepseek config set completion_model deepseek-coder
deepseek config set chat_stream_style markdown
deepseek config set tavily_api_key tvly-live-xxxxxxx
deepseek config unset model

If the config directory is unwritable, fall back to environment variables.

Programmatic configuration

The published module name is deepseek_agent, which now exposes helpers for configuring API keys without touching global environment variables:

from deepseek_agent import configure_api_keys, get_deepseek_api_key

configure_api_keys(
    deepseek_api_key="ds-live-xxxx",
    tavily_api_key="tvly-live-xxxx",
)
print(get_deepseek_api_key())

The module loads DEEPSEEK_API_KEY and TAVILY_API_KEY from the environment by default but lets you override them per process using set_deepseek_api_key and set_tavily_api_key.

Usage

Interactive agent (default)

Running deepseek with no arguments launches the interactive coding agent, now presented through a colourful Rich-powered shell. A command palette is displayed on start so you can see the available /, @, or : shortcuts at a glance (for example @workspace, @model, @read-only, @transcript, @help, @api, and @tavily). Exit with @quit or Ctrl+C. Each request runs as soon as you press Enter—include follow-up guidance in your initial prompt. The assistant appends internal follow-ups that run automated tests and regression checks until they succeed or a clear justification is provided. Use @global on when you need to edit files outside the active workspace. During execution the shell streams the agent's thought process (▌ lines) while non-shell tools render as bright spinners with elapsed time, mirroring modern coding CLIs. Tool outputs are still truncated if they exceed the configured limits; narrow the scope or request additional reads for more detail.

Layer Claude Code-inspired guidance directly inside the shell: @workflow list previews the bundled workflows (feature development or PR reviews), @workflow feature-dev enables the seven-phase plan, and @workflow none clears it. Likewise, @persona code-architect/code-explorer/code-reviewer switch between specialist voices on demand. The same overlays can be preselected for any run via --workflow … and --persona ….

If no API key is detected, the CLI now prompts you to paste one on launch and safely stores it. You can update the stored key at any time with @api or via deepseek config set api_key.

Auto-router smart search chat

deepseek autorouter launches a lightweight chat loop that prefers deepseek-reasoner, streams SSE output (including thinking/status events when available), and heuristically runs Tavily smart search when the query looks time-sensitive. Use --force-search or --no-search to override the heuristic, adjust Tavily depth/result counts, and pass --verbose-router to see the per-turn search decision. When search runs, the numbered context is fed into the model and reprinted beneath the response so you know which sources were cited.

Tavily utilities

Need up-to-the-minute docs without spinning up the full agent? The new deepseek tavily subcommand exposes Tavily's APIs directly:

deepseek tavily search "infra as code adoption" --search-depth advanced --max-results 8
deepseek tavily extract https://example.com/spec --extract-depth advanced --max-pages 3

Both commands automatically use TAVILY_API_KEY (or the stored config value) and share the bundled developer key as a fallback. The agent itself now has a tavily_extract tool alongside tavily_search, so workflows and personas can pull in documentation snippets mid-run without leaving the shell.

Verify installation

deepseek --version
# or use the legacy alias if preferred
deepseek-cli --version

Get help for the CLI:

deepseek --help

One-off agent run

Invoke a single task without entering the interactive shell:

deepseek --prompt "Refactor the HTTP client" \
  --workspace ~/code/project --max-steps 30 \
  --transcript transcript.jsonl --no-global

• Combine --follow-up "Also add tests" to append extra instructions.
• Pass --read-only to prevent write operations and --quiet to suppress verbose progress logs.

Inline shortcuts

Inside the interactive shell, use the new @ commands for quick tasks:

• @chat "Summarise the last commit" streams a single chat response using the current session model and system prompt.
• @complete "def fib(n):" issues a Codex-style completion (results print immediately in the shell).
• @embed "vectorize me" "and me" generates embeddings for one or more snippets.
• @models --filter coder --limit 5 lists available models, with --json for raw output. All other //: controls (@workspace, @model, @transcript, etc.) continue to work as before.

Transcripts and workspaces

• Relative transcript paths under agent mode are resolved within the selected workspace.
• Chat transcripts default to ~/.config/deepseek-cli/transcripts/ when a file name (not path) is supplied.

Legacy shim

Running python deepseek_agentic_cli.py prints a compatibility notice and forwards the call to deepseek --prompt …, so existing automation keeps working.

Publishing to PyPI

1. Update the version in pyproject.toml and commit your changes.
2. Remove old build artifacts:

   rm -rf build dist *.egg-info
   

3. Install packaging tooling:

   python -m pip install --upgrade build twine
   

4. Build the source and wheel distributions:

   python -m build
   

5. Verify the archives:

   python -m twine check dist/*
   

6. Upload to TestPyPI (optional but recommended):

   python -m twine upload --repository testpypi dist/*
   

7. Upload to PyPI:

   python -m twine upload dist/*
   

After publishing, users can install with pip install deepseek-agent.

Development

• python -m deepseek_cli --version exercises the module entry point.
• python -m deepseek_cli --help displays the unified CLI options.
• python -m deepseek_cli config --help shows configuration helpers.
• Run ruff, pytest, or other tooling as required by your workflow.

Contributions welcome! Open issues or pull requests to extend functionality.