Skip to content

Ascendral/codebot-ai

Repository files navigation

CodeBot AI

Autonomous code execution with built-in governance.

The only AI coding agent that runs locally, works with any LLM, and enforces safety policies on every action it takes.

npm version npm downloads license tests node GitHub stars last commit

The Problem

AI coding agents can write code, run commands, browse the web, and modify your filesystem. Most of them do this with zero oversight — no audit trail, no policy enforcement, no way to prove what happened or why.

That's fine for side projects. It's a dealbreaker for teams shipping production software, companies in regulated industries, and anyone who needs to answer the question: what did the AI actually do?

What CodeBot Does Differently

Governance-first architecture. Every tool call passes through a constitutional safety engine (CORD) that risk-scores actions across 14 dimensions before execution. Every action is logged to a SHA-256 hash-chained audit trail with SARIF export. Destructive operations require explicit approval. The agent can't bypass its own safety layer.

Any LLM, anywhere. Run fully local with Ollama, LM Studio, or vLLM — zero API keys, code never leaves your machine. Or connect to Anthropic, OpenAI, Google, DeepSeek, Groq, Mistral, or xAI. Switch models mid-session.

Zero external dependencies. The entire agent runtime — HTTP server, policy engine, audit system, tool execution, provider abstraction — is built on Node.js built-ins. No Express, no Axios, no ORM. The dependency tree is the codebase.

npm install -g codebot-ai
codebot --setup                    # auto-detects local and cloud LLMs
codebot "refactor auth to use JWT" # single task
codebot --dashboard                # web UI at localhost:3120

Who This Is For

  • Security-sensitive development teams — code never leaves your machine with local LLMs
  • Regulated industries (finance, healthcare, government) — audit trail meets compliance requirements
  • Teams that can't send code to third-party APIs — fully self-hosted, no cloud dependency
  • Solo developers who want autonomous coding without subscription fees
  • Internal platform teams building custom dev tooling on top of an extensible agent

Architecture

User ──> Agent Loop ──> Tool Router ──> CORD Safety Engine ──> Execution
              │              │                │                    │
              │              │           14-dimension          32 tools
              │              │           risk scoring          (code, shell,
              │              │           + policy gates         browser, git,
              │              │           + audit trail          network, DB...)
              │              │                │
              │         Permission Model      │
              │         (auto/prompt/         Audit Logger
              │          always-ask)     (hash-chained, SARIF)
              │
         8 LLM Providers
         (local + cloud)
Layer Implementation
Constitutional Safety (CORD) 14 risk dimensions, 5-phase evaluation pipeline, hard-block rules for injection/drift/impersonation
Audit Trail SHA-256 hash-chained logs, tamper-evident, SARIF export for CI integration
Policy Engine Declarative JSON rules — scope file paths, block network targets, restrict commands
Sandbox Docker-based execution with network/CPU/memory limits
Secret Detection 15+ patterns (AWS keys, tokens, private keys) blocked before exposure
SSRF Protection Blocks localhost, private IPs, cloud metadata endpoints

Capabilities

32 built-in tools covering code editing, shell execution, Chrome automation (CDP), web search, deep research, Git operations, Docker management, database queries, SSH, scheduled routines, and persistent memory.

10 app connectors — GitHub, Jira, Linear, Slack, Gmail, Google Calendar, Notion, Google Drive, OpenAI Images, Replicate. Credentials stored in AES-256-GCM encrypted vault.

Web dashboard at localhost:3120 — sessions, audit trail, metrics, security feed, command center.

All 32 tools
Tool Permission Description
read_file auto Read files with line numbers
write_file prompt Create or overwrite files (undo snapshots)
edit_file prompt Find-and-replace edits with diff preview
batch_edit prompt Multi-file atomic find-and-replace
execute always-ask Run shell commands (security-filtered)
glob auto Find files by pattern
grep auto Search file contents with regex
git prompt Git operations (status, diff, log, commit, branch)
browser prompt Chrome automation via CDP
web_fetch prompt HTTP requests and API calls
web_search prompt Internet search with summaries
deep_research prompt Multi-source research with synthesis
think auto Internal reasoning scratchpad
memory auto Persistent memory across sessions
routine prompt Schedule recurring tasks with cron
code_analysis auto Symbol extraction, imports, outline
code_review auto Security scanning and complexity analysis
multi_search auto Fuzzy search: filenames, content, symbols
task_planner auto Hierarchical task tracking
diff_viewer auto File comparison and git diffs
test_runner prompt Auto-detect and run tests (jest, vitest, pytest, go, cargo)
docker prompt Container management (ps, run, build, compose)
database prompt Query SQLite databases (blocks destructive SQL)
http_client prompt Advanced HTTP with auth and headers
image_info auto Image dimensions and metadata
pdf_extract auto Extract text and metadata from PDFs
ssh_remote always-ask Remote command execution via SSH
notification prompt Webhook notifications (Slack, Discord)
package_manager prompt Dependency management (npm, yarn, pip, cargo, go)
app_connector prompt 10 app integrations via encrypted vault
graphics prompt Image processing: resize, crop, watermark, convert

auto = runs silently. prompt = asks first (skipped in --autonomous). always-ask = always confirms.

8 LLM providers
Provider Models
Local (Ollama/LM Studio/vLLM) qwen2.5-coder, qwen3, deepseek-coder, llama3.x, mistral, phi-4, codellama, starcoder2
Anthropic claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5
OpenAI gpt-4o, gpt-4.1, o1, o3, o4-mini
Google gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash
DeepSeek deepseek-chat, deepseek-reasoner
Groq llama-3.3-70b, mixtral-8x7b
Mistral mistral-large, codestral
xAI grok-3, grok-3-mini

Cloud providers route through OpenAI-compatible APIs. Local providers connect directly.

Comparison

CodeBot AI GitHub Copilot Cursor Claude Code
Self-hosted / local LLM Yes No No No
Any LLM provider 8 providers GPT only Mixed Claude only
Constitutional safety engine Yes No No No
Cryptographic audit trail Yes No No No
Sandboxed execution Yes No No No
Free / MIT Yes $10-39/mo $20/mo $20/mo

Extensibility

Programmatic API:

import { Agent, AnthropicProvider } from 'codebot-ai';

const agent = new Agent({
  provider: new AnthropicProvider({
    apiKey: process.env.ANTHROPIC_API_KEY,
    model: 'claude-sonnet-4-6',
  }),
  model: 'claude-sonnet-4-6',
  autoApprove: true,
});

for await (const event of agent.run('list all TypeScript files')) {
  if (event.type === 'text') process.stdout.write(event.text || '');
}

Custom tools — drop .js files in .codebot/plugins/. MCP servers — configure in .codebot/mcp.json. VS Code extension — sidebar chat, inline diffs, status bar. GitHub Actionuses: Ascendral/codebot-ai/actions/codebot@v2

Testing

1,413 tests across 242 suites. CI runs on 3 OS (macOS, Linux, Windows) x 3 Node versions (18, 20, 22). Zero network calls in the test suite — all providers mocked. Security tests cover SSRF, path traversal, injection, and secret detection.

npm test

Build from Source

git clone https://github.com/Ascendral/codebot-ai.git
cd codebot-ai
npm install && npm run build
./bin/codebot

npm · GitHub · Security · Changelog · Docs

MIT — Ascendral