B-AI Code
A coding agent for VS Code that runs on a local model (Ollama) or a frontier API. The harness stays the same;
the model is a setting.
Features
Models
- Works with local models through Ollama, Claude through the Anthropic API, and any OpenAI-compatible endpoint.
- Finds Ollama on your PC by itself and lists the models that can call tools.
- Switch model mid-chat: the conversation continues with the same tools and rules.
- Context size and thinking mode are set automatically for each model.
Coding
- Reads, searches and edits files in your workspace, then runs your build and tests to check its work.
- Every changed file has a diff button, and Undo all restores everything.
- Reads project instructions from
BALA-AI.md or AGENTS.md at the repository root.
Safety
- Every shell command asks for approval, except prefixes you allowlist (such as
npm test).
- Dangerous commands are always blocked:
sudo, rm -rf /, force-push, git reset --hard, curl … | bash.
.git/, .env files and the agent's own policy files are protected. Nothing outside the workspace is touched.
- Per-repository rules in
.bala-ai/policy.json: blocked commands, protected paths, auto-approved commands.
Built for small context windows
- Skills: instructions in a
SKILL.md that load only when a task needs them. Skills written for Claude Code
and Codex (.claude/skills, .agents/skills) work unchanged.
- Helper: a read-only second agent with its own context. It digs through large logs or many files and returns
only a short answer.
- Compaction: older tool output is condensed automatically before the window fills.
- Loop detection: repeating the same call three times triggers a change of approach.
Tools (MCP)
- Connects to any MCP server over stdio or streamable HTTP: GitHub, issue
trackers, mail, documentation.
- Reads
.bala-ai/mcp.json and the .mcp.json that Claude Code uses.
- Tool schemas load on demand, so many servers fit in a 32K window.
- Every MCP call asks for approval unless allowlisted. Tokens stay in environment variables, never in logs.
Privacy and audit
- No telemetry. Prompts and code go only to the model you pick.
- API keys are kept in VS Code's secret storage.
- Every request, tool call and token count is logged to
.bala-ai/logs/.
Benchmark
- A built-in exam compares harnesses on the same local model: 23 embedded-C tasks and 12 web-development tasks,
graded by hidden tests in a sandbox.
- Runs B-AI against Claude Code, Codex CLI, OpenCode, Pi, Qwen Code, Hermes, DeepSeek Harness, Aider and
Harness², and builds a results page with confidence intervals.
Get started
- Install B-AI Code from the VS Code Marketplace, or search for it in the Extensions view.
- Open a folder and click the B-AI icon in the sidebar.
- Pick a model in the dropdown. For Claude, it asks for your API key the first time.
The panel has three tabs: Chat, How to (a short guide) and Settings. Every setting is optional.
Add an MCP server
Run B-AI: Add MCP Server, or write .bala-ai/mcp.json:
{
"mcpServers": {
"github": {
"url": "https://api.githubcopilot.com/mcp/",
"headers": { "Authorization": "Bearer ${env:GITHUB_TOKEN}" },
"description": "GitHub issues, pull requests and code search"
}
}
}
License
Free to install and use. The source code is private and not licensed for reuse. See the licence file included with the extension.
| |