vulpcode 0.2.0

Terminal coding agent, multi-provider (Anthropic, OpenAI, Gemini, Ollama, ...)

Vulpcode

A terminal coding agent that lets you choose the model.

Vulpcode is a CLI agent for software engineering tasks — inspired by Claude Code — that is provider-agnostic. The same set of tools, slash commands, and MCP support, but you decide which model to talk to: paid APIs (Anthropic, OpenAI, Gemini), local engines (Ollama, LM Studio, vLLM), OpenAI-compatible gateways (DeepSeek, Groq, OpenRouter), or internal corporate endpoints.

---

Highlights

• Provider-agnostic. Switch models with a single flag (--provider, --model), without changing your workflow.
• Privacy-first. Run fully offline with Ollama, LM Studio, or vLLM. No outbound traffic when the chosen provider is local.
• Pip-native. Installs like any Python package. No Node, no npm, no external runtime.
• Tool-complete. Functional parity with the leading agentic CLIs: Bash (foreground + background), Read, Write, Edit, MultiEdit, Glob, Grep, WebFetch, WebSearch, Task, TodoWrite, NotebookEdit.
• MCP support. Connects to any Model Context Protocol server you configure; their tools are auto-registered into the agent.
• Hackable. ~3k lines of core code, tools as plugins, providers as adapters.

---

Installation

Requires Python 3.11+.

pip install vulpcode

Installing inside a virtual environment is recommended:

python -m venv ~/.venv/vulpcode
source ~/.venv/vulpcode/bin/activate
pip install vulpcode

Vulpcode ships two equivalent console scripts: vulp (short) and vulpcode.

Optional extras

Extra	Adds	Use it for
[dev]	pytest, pytest-asyncio, ruff, mypy, respx	Contributors running the test suite
[docs]	mkdocs, mkdocs-material, mkdocstrings[python], pymdown-extensions	Building the documentation site
[search]	duckduckgo-search	Using WebSearch without a Tavily key
[docs-tools]	python-docx, reportlab, pypdf, markdown-it-py, nbformat, PyYAML, sqlparse, pydot	Validators for the Write* family (docx/pdf/notebook/etc.)

pip install "vulpcode[search]"
pip install "vulpcode[dev,docs]"
pip install "vulpcode[docs-tools]"

From source

git clone https://github.com/Vulp-Code/vulpcode.git
cd vulpcode
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,docs,search]"

---

Quick start

# Verify the install
vulp --version
vulp providers

First chat with Claude

export ANTHROPIC_API_KEY=sk-ant-...
vulp --auto "say hi in one word"

Run fully offline with Ollama

ollama pull qwen2.5-coder:7b
vulp --provider ollama --model qwen2.5-coder:7b --auto "explain git rebase"

Use an OpenAI-compatible gateway (e.g. Groq)

export GROQ_API_KEY=gsk_...
vulp --provider groq --model llama-3.3-70b-versatile

Use an internal corporate endpoint

The library ships two providers for internal corporate /chatCompletion endpoints:

Provider	When to use
internal-llm	Plain chat only — no file creation, no shell, no agentic loop
internal-llm-agentic	Full agentic flow via a text-based tool protocol (recommended)

Both read the same env vars:

export INTERNAL_LLM_ENDPOINT="https://internal.example.com/v1/chat"
export INTERNAL_LLM_USER_UUID="00000000-0000-0000-0000-000000000000"

# Chat-only mode
vulp --provider internal-llm

# Full agent (creates files, runs validators, retries on syntax errors)
vulp --provider internal-llm-agentic

---

Supported providers

Provider	Backend type	Tools	Vision	Streaming
anthropic	Anthropic (Claude)	x	x	x
openai	OpenAI (GPT)	x	x	x
gemini	Google (Gemini)	x	x	x
ollama	Local	x	x	x
deepseek	OpenAI-compatible	x		x
groq	OpenAI-compatible	x		x
openrouter	OpenAI-compatible	x		x
lmstudio	Local	x		x
vllm	Local	x		x
internal-llm	Corporate endpoint
internal-llm-agentic	Corporate endpoint (text-protocol tool calling)	x

Run vulp providers to list them at any time.

---

Built-in tools

Tool	Purpose
Read	Read files (text, images, notebooks, PDFs)
Write	Create or overwrite a file
Edit	Exact string replacement
MultiEdit	Multiple edits in a single file, atomically
Bash	Run shell commands
BashOutput	Stream output from a long-running background process
KillBash	Terminate a background process
Glob	Find files by pattern
Grep	Search file contents
WebFetch	Fetch and parse a URL
WebSearch	Search the web (Tavily or DuckDuckGo)
Task	Delegate work to a sub-agent
TodoWrite	Persistent task list inside a session
NotebookEdit	Edit Jupyter notebook cells

Validated write tools

Format-specific writers that parse-and-validate the content before saving and write atomically (temp file + rename). If validation fails the target file is never touched, so the agent can repair the input and retry.

Tool	Format	Validation
WritePy	Python	compile() (syntax check)
WriteIpynb	Notebook	nbformat schema
WriteMd	Markdown	markdown-it-py
WriteDocx	Word	python-docx
WritePdf	PDF	pypdf (or reportlab for generation)
WriteJson	JSON	json.loads
WriteYaml	YAML	PyYAML safe-load
WriteToml	TOML	tomllib
WriteCsv	CSV	csv round-trip
WriteXml	XML	xml.etree
WriteHtml	HTML	html.parser
WriteSh	Shell	bash -n syntax check
WriteSql	SQL	sqlparse
WriteSvg	SVG	xml.etree + root-element check
WriteDot	Graphviz	pydot

Install the non-stdlib validators with pip install "vulpcode[docs-tools]".

---

Slash commands

Available inside the REPL:

Command	What it does
/help	Show all commands
/clear	Clear the screen and conversation history
/exit	Leave the REPL
/tools	List enabled tools
/cost	Show token usage and estimated cost
/compact	Compact the conversation to free context
/provider <name>	Switch provider
/model <id>	Switch model on the current provider
/save <name>	Save the current session
/load <name>	Restore a saved session
/mcp	Inspect connected MCP servers

---

Permission modes

Vulpcode never executes tools silently by default. Choose a mode that matches your trust level for the current task:

Mode	Flag	Behavior
default		Read tools auto-run; writes and Bash require confirmation
auto	--auto	Auto-approve every tool call
safe	--safe	Confirm every tool call, including reads
plan	--plan	Plan-only — the agent proposes actions but executes nothing

Headless mode for scripting:

vulp --print --auto "refactor src/utils.py to remove duplication"

Resume your last session:

vulp --resume

---

MCP (Model Context Protocol)

MCP servers declared in ~/.vulpcode/config.toml are launched automatically. Their tools are added to the registry with the prefix mcp__<server>__<tool> and become available like native tools.

Minimal config:

[mcp.servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/home/me/projects"]

Inspect runtime state with /mcp inside the REPL.

---

Configuration

Vulpcode reads configuration in this order:

1. CLI flags (--provider, --model, ...)
2. Environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY, OLLAMA_HOST, INTERNAL_LLM_ENDPOINT, ...)
3. ~/.vulpcode/config.toml

Sessions, history, and per-project state live under ~/.vulpcode/.

---

Status

Alpha (current release: 0.2.0). The CLI surface, tool set, and config format may still change before 1.0. See CHANGELOG.md for release notes.

---

Documentation

Full documentation, including detailed configuration, recipes, the API reference, and architecture notes, is built with MkDocs from the docs/ directory.

pip install "vulpcode[docs]"
mkdocs serve

---

Contributing

Issues and pull requests are welcome. To set up a development environment:

git clone https://github.com/Vulp-Code/vulpcode.git
cd vulpcode
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,docs,search]"
pytest
ruff check .
mypy

---

License

MIT — see LICENSE.

---

Links

• Source code: https://github.com/Vulp-Code/vulpcode
• Issues: https://github.com/Vulp-Code/vulpcode/issues
• PyPI: https://pypi.org/project/vulpcode/