Anvil

Local AI coding that understands your codebase.

Repository-aware · Local-first · Ollama · Open source

Anvil is a local-first AI coding IDE built on the Code-OSS / VSCodium ecosystem. It explores your repository, traces relevant implementation, and answers from project evidence instead of guessing from filenames. Local models via Ollama are a first-class path.

Anvil IDE — local AI coding with repository-aware assistance.

Developer Preview — Build from source. There is no official prebuilt binary download yet.

---

Quick Start

macOS (packaged app from source)

git clone https://github.com/pranee54/Anvil.git
cd Anvil
npm install

Install Ollama, then:

ollama pull qwen2.5-coder:3b
npm run anvil:package:mac
open dist/Anvil.app

If macOS blocks the app: Finder → right-click Anvil.app → Open. Details: Install macOS.

Then:

1. File → Open Folder… — try demo-project/
2. Select Ollama
3. Select qwen2.5-coder:3b
4. Run Anvil: Test Ollama Connection
5. Open Anvil Chat (Cmd+L)
6. Ask: Explain this project.

Windows (development only)

There is no standalone Windows app yet. Build packages and run Anvil Agent in an Extension Development Host — see Install Windows.

---

Why Anvil?

Repository-aware investigation	Searches and reads relevant project files before answering codebase questions.
Evidence-grounded answers	Responses can reference inspected project files and sources when evidence is available.
Local-first AI	Run supported Ollama models on your machine.
Agent workflows	Ask, Edit, and Agent modes for different development tasks.
Developer control	Tool execution follows Anvil’s permission system (strict / ask / permissive).

Investigation improves grounding; it does not guarantee perfect correctness.

---

Anvil in Action

Ask Mode Repository-aware answers grounded in project files.

Agent Investigation Trace implementation across the repository before answering.

Local AI Run supported Ollama coding models locally.

---

How Anvil Investigates

For codebase questions, Anvil attempts to gather repository evidence before answering:

Evidence before answers — investigation improves grounding; it does not guarantee perfect correctness.

You may see investigation steps such as searching the codebase, reading files, and citing sources. Depth depends on the question, workspace size, model, and permissions.

---

Local AI

Anvil does not ship Ollama or model weights. Install them separately and connect Anvil to your local server (default http://127.0.0.1:11434).

Provider	Status
Ollama	Recommended · local · streaming
OpenAI-compatible HTTP	Implemented · responses are not streamed
Anthropic	Not enabled (scaffolded only)
Gemini	Not enabled (scaffolded only)

Suggested coding models

Tier	Model
Light	qwen2.5-coder:3b
Balanced	qwen2.5-coder:7b
Larger	qwen2.5-coder:14b

Choose based on your machine’s resources. Full setup: Local AI.

---

Modes

Mode	Role
Ask	Understand and explain without intentionally modifying files.
Edit	Focused code changes with review (Accept / Reject).
Agent	Multi-step repository tasks using allowed tools.

Attach context with @file, @selection, @folder, @codebase, @problems, and @git. Details: Usage.

---

Feature Status

Feature	Status
Repository investigation	Available
Evidence-grounded answers	Available
Ask mode	Available
Edit mode	Available
Agent mode	Available
Ollama	Available
OpenAI-compatible endpoint	Available
Anthropic	Not enabled
Gemini	Not enabled
macOS source-built app	Available
Windows development workflow	Available
Windows packaged app	Planned
Signed / notarized macOS distribution	Planned

---

Try Anvil Safely

The repo includes demo-project/ — a small Task Board sample for testing repository investigation without opening private code.

Suggested prompts:

• Explain this project.
• How does authentication work?
• Find the API implementation.
• Find a bug in this project.

---

Platform Support

Platform	What works today
macOS	Build from source → dist/Anvil.app (ad-hoc / unsigned; Gatekeeper may require right-click → Open).
Windows	Extension Development Host only — no standalone download.
Linux	Not verified — planned.

---

Documentation

Guide	Description
Install macOS	Package and launch Anvil.app
Install Windows	Extension Development Host workflow
Local AI	Ollama setup and model tiers
Usage	Modes, @ context, permissions
Troubleshooting	Common failures
Development	Contributor setup
Architecture	System design
Security	Vulnerability reporting
Contributing	How to contribute

---

Architecture

Code-OSS / VSCodium
        ↓
Anvil Extension
        ↓
    agent-core
   ↙    ↓     ↘
Context Tools Models
              ↓
            Ollama

Path	Role
packages/agent-core	Orchestrator, investigation, tools, permissions, models
packages/anvil-extension	Anvil Chat UI, commands, diffs, checkpoints
ide/	Branding + macOS packaging
demo-project/	Safe sample workspace
apps/legacy-electron/	Archived prototype — not the product path

Details: Architecture.

---

Development

npm install
npm run build:agent
npm run build:extension
npm run typecheck:agent
npm run typecheck:extension
npm test

See Development and Contributing.

---

Contributing

Issues and feature requests are welcome via GitHub Issues.

• CONTRIBUTING.md — setup, style, PR expectations
• SECURITY.md — report vulnerabilities privately when possible

Keep Anvil-specific logic in packages/agent-core and packages/anvil-extension.

---

Security

Report vulnerabilities per SECURITY.md. Do not paste API keys, tokens, or private user data into public issues.

---

License

Anvil-owned code: MIT.

Packaged desktop builds redistribute a Code-OSS / VSCodium-derived IDE. See THIRD_PARTY_NOTICES.md.