DevPilot 🚀

An autonomous AI coding agent for your terminal. DevPilot gives Claude, GPT, Groq, Mistral, or any local model a full suite of tools — file read/write, bash execution, code search, git, MCP, and more — orchestrated into a self-healing agentic loop inside a rich Textual TUI. Works in PowerShell, CMD, and any modern terminal.

┌─ DevPilot ──────────────────────────────────────────────────────┐
│  ✦ claude-opus-4-5  │  agent/tools/fs.py  │  iteration 2/50    │
├─────────────────────────────────────────────────────────────────┤
│  You › Add input validation to the write_file tool              │
│                                                                 │
│  ● Calling read_file   agent/tools/fs.py                        │
│  ● Calling edit_file   agent/tools/fs.py                        │
│    ✓ Edited agent/tools/fs.py successfully.                     │
│                                                                 │
│  Done — added a size cap and path-escape guard. The diff is     │
│  above. Want me to run the tests?                               │
└─────────────────────────────────────────────────────────────────┘

---

Features

Category	Capability
File ops	read_file, write_file, edit_file (surgical replace), list_files
Shell	run_bash with timeout, blocked-command guard, and auto-heal loop
Code search	Regex search_code
Pre-flight linting	Syntax-checks Python (ast.parse) and JS/TS (node --check) before writing
Git	git_status, git_commit with surgical file staging
Documentation	doc_gen (markdown), diagram (Mermaid)
Web search	web_search via Tavily (optional)
MCP	Connect any MCP server via mcp_servers.json
A2A	Agent-to-agent task delegation over HTTP
Providers	Anthropic, OpenAI, Groq, Together AI, Mistral, Ollama, any OpenAI-compatible endpoint

---

Quickstart

Requires Python ≥ 3.11.

pip install devpilot-agentic-cli
devpilot

That's it. On first run DevPilot launches a setup wizard — pick your provider, enter your API key, choose a model. It saves everything to ~/.devpilot/.env and opens the TUI. Every subsequent run starts immediately.

Optional Features

PDF Documentation Generation Enable the generate_docs tool to export PDFs:

pip install devpilot-agentic-cli[pdf]

Note: You also need to install the wkhtmltopdf system binary.

---

Providers

DevPilot works with any of these out of the box — the setup wizard walks you through each one:

Provider	Models	Get API key
Anthropic	claude-opus-4-7, claude-opus-4-5-20251101, claude-sonnet-4-5-20250929, claude-haiku-4-5-20251001	console.anthropic.com
OpenAI	gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano	platform.openai.com
Groq	llama-3.3-70b-versatile, llama-4-scout-17b, qwen3-32b	console.groq.com
Together AI	Llama 4 Maverick/Scout, Llama 3.3 70B Turbo	api.together.xyz
Mistral AI	mistral-large-latest, mistral-small-latest, codestral-latest	console.mistral.ai
Ollama	qwen2.5-coder:7b, deepseek-coder-v2:16b, llama3.3:70b — no API key needed	ollama.com
Other	Any OpenAI-compatible endpoint	—

---

Configuration

All settings live in ~/.devpilot/.env — a persistent file in your home directory that works regardless of which folder you run devpilot from. The setup wizard creates this on first run. You can edit it manually anytime — or re-run the wizard with devpilot --setup.

Full settings reference

Variable	Default	Description
ANTHROPIC_API_KEY	—	API key for Anthropic provider
OPENAI_API_KEY	—	API key for OpenAI / Groq / Together / Mistral / Ollama
GROQ_API_KEY	—	API key for Groq (optional — can reuse as OPENAI_API_KEY)
DEVPILOT_PROVIDER	anthropic	anthropic or openai
DEVPILOT_MODEL	claude-opus-4-5-20251101	Model name
DEVPILOT_BASE_URL	—	Custom endpoint, e.g. https://api.groq.com/openai/v1
DEVPILOT_NO_CONFIRM	false	Skip confirmation prompts (useful for CI)
DEVPILOT_MAX_ITERATIONS	50	Max tool-use iterations before loop aborts
DEVPILOT_WORKDIR	cwd	Root directory for file operations
DEVPILOT_SESSIONS_DIR	.devpilot_sessions	Where session JSON files are saved
DEVPILOT_THINKING	false	Enable extended thinking (Claude only)
DEVPILOT_THINKING_BUDGET	10000	Token budget for extended thinking

Override priority

CLI flags  →  env vars  →  ~/.devpilot/.env  →  defaults

Per-run override example:

DEVPILOT_MODEL=llama-3.3-70b-versatile devpilot --task "fix typos"

---

CLI Reference

devpilot [OPTIONS]

Options:
  --provider {anthropic,openai}   Model provider
  --model MODEL                   Model name
  --base-url URL                  OpenAI-compatible base URL (e.g. Ollama)
  --no-confirm                    Skip confirmation prompts
  --thinking                      Enable extended thinking (Claude only)
  --thinking-budget N             Token budget for extended thinking (default: 10000)
  --workdir PATH                  Working directory for file operations
  --task TASK                     Run a single task and exit (CI mode)
  --resume SESSION_ID             Resume a previous session
  --a2a-port PORT                 A2A server port (default: 8000)
  --no-a2a                        Disable A2A server
  --no-web-search                 Disable Tavily web search
  --setup                         Re-run the setup wizard

CI / single-task mode

devpilot --task "Run the test suite and fix any failures" --no-confirm

---

Using Ollama (Local Models)

# Pull and serve a model
ollama pull qwen2.5-coder:7b
ollama serve

# Run devpilot and pick Ollama in the setup wizard
devpilot --setup

Or set env vars directly:

DEVPILOT_PROVIDER=openai \
DEVPILOT_BASE_URL=http://localhost:11434/v1 \
DEVPILOT_MODEL=qwen2.5-coder:7b \
devpilot

---

MCP Integration

DevPilot connects to any Model Context Protocol server. Edit mcp_servers.json in your project root:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/dir"]
    }
  }
}

Tools from connected MCP servers are automatically available to the agent on startup.

---

Architecture

agent/
├── cli.py                Entry point — args, setup wizard, wires everything
├── setup_wizard.py       First-run interactive configuration wizard
├── loop.py               Core agentic loop (plan → act → verify → heal)
├── config.py             Config dataclass — all settings from env vars
├── context.py            RepoContext — file awareness, AST project map
├── history.py            Conversation history + smart context pruning
├── providers/
│   ├── anthropic_provider.py
│   ├── openai_provider.py
│   ├── system_prompt.py  PLAN→ACT→VERIFY prompt, platform-aware shell rules
│   └── factory.py
├── tools/
│   ├── fs.py             read_file, write_file (pre-flight lint), edit_file, list_files
│   ├── shell.py          run_bash
│   ├── search_code.py    regex search
│   ├── git_ops.py        git_status, git_commit
│   ├── doc_gen.py        doc_gen
│   ├── diagram.py        diagram (Mermaid)
│   ├── web_search.py     web_search (Tavily)
│   ├── a2a.py            A2A delegation
│   └── registry.py       ToolRegistry + PermissionGuard
└── tui/
    └── app.py            Textual TUI — chat log, project map, tool drawer

---

Running Tests

pip install devpilot-agentic-cli[dev]
pytest

48 tests — unit tests for every tool plus end-to-end integration tests with a mocked provider.

---

Development / Contributing

git clone https://github.com/Thijeshpraveen-V/DevPilot
cd DevPilot
pip install -e ".[dev]"
devpilot --setup

PR checklist:

• pytest passes (48 tests, 0 failures)
• No new hard dependencies without discussion
• New tools follow the BaseTool pattern in agent/tools/base.py

---

License

MIT © Thijesh Praveen V