KendaliAI (Go Edition)

KendaliAI is a self-hosted, autonomous AI coding agent and orchestration gateway rebuilt natively in Go. Influenced heavily by the "Claude Code in 200 Lines" architecture, KendaliAI uses dynamic tool invocation and recursive LLM cognition loops to navigate, evaluate, and edit your local filesystem.

Features

• Autonomous Cognition Loop: Recursive OS-level operations (read_file, list_files, edit_file, bash) executed dynamically by models like DeepSeek.
• OpenAI-Compatible Provider: Works with any OpenAI-compatible endpoint (DeepSeek, OpenAI, Groq, Ollama, etc.) — just change config.json.
• Custom Skills System: Extend the agent's capabilities with on-demand Markdown instructions or custom shell-based tools.
• Dynamic Configured Persona: Your AI's identity and restricted commands are defined dynamically in ~/.kendaliai/Persona.md.
• Local TUI Dashboard: An interactive Terminal User Interface using BubbleTea for direct agent interaction.
• Telegram Gateway: Execute local terminal commands securely from your phone.
• Centralized Telemetry Logging: Stream OS Agent logs from all channels seamlessly.

Quickstart

Prerequisites

• Go 1.24+
• CGO (required by go-sqlite3)

1. Configure

All configuration lives in config.json. Copy the example and fill in your credentials:

cp config.example.json config.json

{
  "chatProvider": {
    "type": "deepseek",
    "endpoint": "https://api.deepseek.com/v1",
    "model": "deepseek-v4-flash",
    "apiKey": "your-api-key"
  },
  "embedding": {
    "endpoint": "https://api.openai.com/v1",
    "model": "text-embedding-3-small",
    "apiKey": "your-embedding-api-key"
  },
  "channel": {
    "type": "telegram",
    "botToken": "your-telegram-bot-token"
  }
}

Config resolution order:

1. KENDALIAI_CONFIG env var (explicit override)
2. ./config.json (development — local file)
3. ~/.kendaliai/config.json (production — home directory)

channel is optional. Without it, only TUI mode works. When you run onboard or gateway, you'll see instructions on how to add a channel.

2. Install Dependencies

make tidy

3. Onboard

Creates the gateway database and auto-binds the channel from config:

make dev-onboard

Running

Development

Uses ./config.json automatically (no install needed):

# Run TUI dashboard
make dev-tui

# Run headless gateway (Telegram bot + API server)
make dev-gateway

# Run any command
make dev-<command>

Production

# Build optimized binary
make build-prod

# Install binary + config to system
make install

# Run
kendaliai tui
kendaliai gateway

make install copies the binary to /usr/local/bin/ and config.json to ~/.kendaliai/config.json (if not already present).

Makefile Targets

Target	Description
make build	Build debug binary to ./build/
make build-prod	Build optimized binary (stripped, no symbols)
make install	Build + install to /usr/local/bin/ + copy config
make dev	Run via go run (no build step)
make dev-tui	Run TUI dashboard
make dev-gateway	Run headless gateway
make dev-onboard	Initialize gateway + bind channel from config
make dev-logs	Run log streamer
make tidy	Run go mod tidy
make lint	Run go vet
make clean	Remove build artifacts

Architecture

KendaliAI operates in natively decoupled environments. You can run one or multiple components completely asynchronously.

Standalone Interactive TUI (Offline Agent)

Access the autonomous agent locally through a BubbleTea interface. Fully actionable terminal environment with live streaming output.

make dev-tui        # development
kendaliai tui       # production

Headless Gateway (Telegram Bot)

Starts the primary server and polls attached Telegram bots.

make dev-gateway    # development
kendaliai gateway   # production

Centralized Logistics Stream

Watch the autonomous agent think, execute tools, and respond in real-time across the entire platform.

make dev-logs       # development
kendaliai logs      # production

Custom Skills

KendaliAI supports two types of custom skills located in ~/.kendaliai/skills/.

1. Instructional Skills (.md)

Add specialized knowledge or guidelines by creating a Markdown file with YAML frontmatter. These are registered as "on-demand" tools that the agent calls to load expert instructions when needed.

Example: ~/.kendaliai/skills/frontend-design.md

---
name: frontend-design
description: Create distinctive, production-grade frontend interfaces.
---
## Principles
- Use modern typography (Outfit, Roboto).
- Avoid generic "AI slop" aesthetics.
- Prioritize CSS-only motion effects.

2. Execution Skills (.sh + skills.json)

Add functional tools that execute shell scripts. These must be defined in ~/.kendaliai/skills/skills.json.

Example: ~/.kendaliai/skills/weather.sh

#!/bin/bash
curl -s "wttr.in/$1?format=3"

Registering in skills.json:

{
  "skills": [
    {
      "id": "weather",
      "name": "Get Weather",
      "description": "Fetch current weather for a city.",
      "input_schema": {
        "type": "object",
        "properties": { "city": { "type": "string" } }
      },
      "execution": {
        "type": "shell",
        "command": "./weather.sh",
        "args_mapping": { "city": "$1" }
      },
      "installed": true
    }
  ]
}

System Structure

cmd/kendaliai/       # Primary CLI entrypoints (root, tui, gateway, logs)
internal/
├── agent/           # The Core Cognition Loop & Native Tool Registry
├── channels/        # External polling wrappers (Telegram)
├── config/          # Config singleton (config.json)
├── db/              # SQLite workspace storage
├── gateways/        # State handlers
├── logger/          # Central syslog mapping
├── providers/       # OpenAI-compatible provider (any endpoint)
├── security/        # Identity security
├── server/          # REST Gateway wrappers
└── tui/             # Charmbracelet Bubbletea reactive loop

Security & Restricting Commands

Your agent is strictly bounded to the rules specified inside ~/.kendaliai/Persona.md. If this file doesn't exist, KendaliAI will generate it upon execution.

To restrict commands from being blindly executed natively by the agent, define them under exclude_cmd: in the file:

# Agent Identity

**Name:** KendaliAI

tools: read_file, list_files, edit_file, bash
exclude_cmd: rm, ls ., modify root file

Switching Providers

KendaliAI uses a single OpenAI-compatible provider. To switch models or providers, just edit config.json:

{
  "chatProvider": {
    "type": "openai",
    "endpoint": "https://api.openai.com/v1",
    "model": "gpt-4o",
    "apiKey": "sk-..."
  }
}

Works with any OpenAI-compatible API: DeepSeek, OpenAI, Groq, Together, Ollama, LM Studio, etc.