autokeren 0.12.5

Cloudflare-first agentic coding CLI untuk developer Indonesia dan global

autokeren

Cloudflare-first agentic coding CLI with an interactive TUI for developers worldwide.

English | Bahasa Indonesia | 简体中文 | Português (Brasil) | Español | 日本語

autokeren is an agentic coding CLI for the Cloudflare-first stack. Its active runtime and interactive Text User Interface (TUI) are native Go, with Python retained only as an explicit compatibility runtime. It supports automatic model fallback and includes tools for files, shell commands, git, Cloudflare deployments, and the built-in PaaS.

Build Week 2026 Extension — Autokeren Proof

Autokeren existed before OpenAI Build Week (began on 2026-07-04) and is distributed under the MIT license. Autokeren Proof was built during the submission period with Codex and GPT-5.6. It adds an evidence-led release workflow: acceptance criteria, recorded verification, and a SHIP/BLOCKED/NEEDS_HUMAN_REVIEW release verdict.

🚀 Quickstart for Judges (No API Key Required)

Judges can test the proof-rendering system and run the local verification suite without any external API keys or configuration:

1. Install Autokeren & Clone Repository:

   pipx install autokeren
   git clone https://github.com/autokeren/autokeren
   cd autokeren
   

   Supported platforms: Linux, macOS (Intel & Apple Silicon), and Windows (via PowerShell or WSL). Requires Python 3.11+.

2. Run Local Verification Checks: Execute the deterministic demo app suite (validates a local HTTP server with successful and bad requests):

   python3 examples/proof-demo/test_app.py
   

3. Replay the Proof Artifact: Render the visual Release Card for the demo run in your terminal:

   autokeren --proof replay examples/proof-demo/proof-run.json
   

(Note: The replay command displays a pre-recorded, verified proof run containing the test evidence. It does not run the tests dynamically itself.)

🛠️ Hackathon Contribution (What was built)

• OpenAI Model Integration (openai.py): Native model client supporting GPT-5.6 and Codex completions, streaming, and tool calls.
• Native Go Runtime Migration (cmd, internal): Moved the active provider loop, tool dispatch, session/context handling, proof CLI, and provider-aware bootstrap to Go.
• Provider-Aware Login Wizard (bootstrap.go): Go-native setup that validates providers and saves a matching primary and fallback model for the native Go runtime.
• Proof Tool (proof.go): Native Go tool to plan, record, report, replay, and list release evidence runs.
• /proof Slash Command (cli.py, tui.py): Interactive commands to execute proof tasks.
• Deterministic Demo App (app.py): A mock checkout app showing defect validation and verification logs.

🧠 Collaboration Disclosure (Human vs. AI)

• Codex Acceleration: Codex accelerated the Go migration, including the provider router, tool dispatch, durable sessions, proof replay, login bootstrap, and compatibility launcher.
• GPT-5.6 Contribution: GPT-5.6 generated comprehensive unit tests, security validations (path traversal checks), and the strict JSON schema verification engine.
• Human Decisions:
  • Architectural Choice: Moving the IPC channel to a local TCP socket rather than standard output (stdout) to avoid parser crashes caused by dependency prints/warnings.
  • Product Design: The linear state transitions (PLAN -> RECORD -> REPORT -> REPLAY) for Release Cards.
  • Antigravity Decision: The Google Antigravity option was hidden from the interactive wizard to keep the interface clean for public users, while preserving CLI flags (--agy) and configuration support so existing installations do not regress.

🎥 Demo Video & Devpost Submission

• Demo Video: Watch the 3-minute video on YouTube
• Codex Session ID: 019f6df7-0978-7b71-aaa1-8c9b152e5ead (Paste this ID directly into the Devpost form.)

---

Key Features

• 7 AI models — kimi-code, kimi-2.6, glm-5.2, glm-flash, llama-4-scout, gemma-4, and nemotron with automatic fallback.
• Built-in PaaS — deploy applications to Cloudflare Workers directly from the terminal, with automatic D1 + R2 + AI bindings.
• Multi-Agent Mode & Auto Spawn — Run multiple agents in parallel via /project, or let the primary agent dynamically call sub-agents using the spawn_agent tool.
• MCP Server Support — Integrate third-party external tools via Model Context Protocol (MCP) and manage them with /mcp.
• Input History — Navigate previous command entries using the ↑ / ↓ arrow keys in the terminal.
• Export Chat — Export the entire chat history as a Markdown file using the /export command.
• Streaming Output — Token-by-token response rendering in real-time.
• Permission System — Prompts for confirmation before executing potentially dangerous shell commands or modifying files.
• Cross-Session Memory — Automatically stores and loads project-specific persistent memory upon startup.
• Session Save/Resume (SQLite) — Save conversation state to a transactional local SQLite database (sessions.db) and resume anytime using slash commands or the -r CLI flag.
• Context Tracking + /compact — Monitor context window usage and summarize history automatically or manually.
• AGENTS.md Support — Automatically loads project-specific instructions for the AI agent.
• Markdown Rendering — Rich terminal formatting for headings, tables, and syntax-highlighted code blocks.
• KV/D1/PaaS Tools — Read/write KV pairs, run D1 database queries, and manage projects directly from the agent.
• Tmux Supervisor — Spawn and monitor long-running background agents that survive terminal closure.
• CF Pages/Workers Deploy — Integrated helper tools for building and deploying to Cloudflare.
• File Explorer (F7) — Toggle the folder/file tree on the left TUI panel, click a file to automatically read its content.

Vibe Coding Features (v0.8.0)

9 original features not found in other coding CLIs (Claude Code, Aider, Cursor, Cline):

Time-Travel /rewind

Undo tool calls and restore the codebase to previous checkpoints. Automatically saves a checkpoint after every file write/patch.

/rewind        # undo 1 tool call
/rewind 3      # undo 3 tool calls
/rewind list   # list all available checkpoints

Architecture Guardian

Index project genome (modules, functions, dependencies) and block duplicate functions/modules before they are written.

/genome         # view project genome
/genome rescan  # rescan project genome
/genome check   # check for duplicate functions

Loop Breaker

Detect when the agent is stuck in an error/apology/file-thrashing loop. Automatically swaps the active AI model.

/loop status    # view loop breaker error history
/loop reset     # reset loop tracker
/loop break     # manual break — swap model + reset

Cross-Model Auto-Review

Review unstaged or staged diffs using an AI model from a different vendor to catch blind spots.

/review         # review unstaged diffs
/review staged  # review staged diffs

Vibe-Security Guard

Automatically scan every file write for secrets, SQL injection, XSS, and forbidden patterns.

/security           # scan the entire project
/security app.py    # scan a specific file

Live Architecture Enforcement

Rules-based enforcement via .ak-rules.yaml (e.g., maximum file lines, forbidden patterns, import limits).

Spec-Driven Auto-Planning

AI-guided user interview with 20 questions to automatically generate plan.md and technical-plan.md.

/spec build a REST API     # start interview
/spec answer my answer     # answer interview questions
/spec generate              # generate implementation plan
/spec show                  # view implementation plan
/spec progress              # track implementation progress

Ghost Agent

Spawn background agents in tmux for parallel task execution.

/ghost fix bug in login.py  # spawn ghost agent
/ghost list                 # list all background agents
/ghost show 1               # view output of ghost #1
/ghost kill 1               # kill ghost #1
/ghost kill all             # kill all background agents

Research Tool

Deep web search querying Reddit, Hacker News, and Google Web search.

/research python coding tools     # search all sources
/research reddit asyncio tips     # search Reddit only
/research hn AI coding CLI        # search Hacker News only
/research web best practices      # search Google Web only

AGI Evolution & Self-Healing (v0.11.0+)

Bringing autonomous artificial intelligence to Autokeren CLI:

1. Continuous Lifelong Daemon & Observer

An asynchronous background system observer tailing critical error logs and file changes to trigger self-healing processes.

2. Self-Evolution / Auto-Refactoring Loop

Automatically refactors broken Python tools, validates them with new pytest unit tests, and hot-reloads the tool registry.

3. Local TF-IDF Semantic Memory

High-performance local semantic search using Vector Space Model (VSM) with TF-IDF weighting and Cosine Similarity in a local SQLite database (memory.db). No API keys required!

4. Interactive Kanban TUI Board (Ctrl+K)

Manage project task lists visually inside the terminal, synchronized with local SQLite. Press Ctrl+K to toggle the board at any time.

5. Live Multi-Agent Debate View (Ctrl+D)

Monitor discussions, coordination, and work logs of background Ghost Agents in real-time. Press Ctrl+D to toggle the debate view.

6. AST Dependency Graphing (Poka-Yoke)

Recursively parses imports in Python, Go, and JS/TS files to construct a local dependency graph. Feeds the agent with contextually relevant signatures, preventing compile errors.

7. CAPA Tag-Based Rollback

Creates tag checkpoints (ak-green-) after successful micro-commits, and automatically rolls back to the latest verified green state upon loop breaker failure, injecting warning history context to prevent error recursion.

8. Active Tmux Log Sniffing

Asynchronously checks tmux sessions for traceback errors, panics, and server crashes, automatically triggering background self-healing processes.

Installation & Setup

1. Get a Free API Key

Sign up at developers.autokeren.com to get your API key.

2. Install

Linux / macOS

pipx install autokeren

If you don't have pipx: sudo apt install pipx && pipx ensurepath (Linux) or brew install pipx (macOS) Alternative: pip install --user autokeren

Windows (PowerShell)

Step 1 — Install pipx via pip:

python -m pip install --user pipx

Step 2 — Add pipx to Windows PATH:

python -m pipx ensurepath

Step 3 — Restart PowerShell (close and reopen the terminal).

Step 4 — Install autokeren:

pipx install autokeren

3. Login

autokeren --login

Enter your API key from developers.autokeren.com.

4. Start Coding

autokeren

Quick Start

Interactive TUI Chat (Default)

Launch the interactive TUI interface:

autokeren

Single Prompt (Non-interactive)

autokeren "create a hello.py file that prints hello world"

Plan Mode

autokeren --plan

Resume Saved Session

Resume a saved session directly from the terminal startup:

autokeren --resume my-session-name
# or using the short flag
autokeren -r my-session-name

Choose Model

autokeren -m glm "refactor this function"
autokeren -m kimi "write unit tests for the tools module"

Google AI Studio Mode (Gemini API)

Run autokeren directly using your own Google AI Studio API key:

autokeren --aistudio

If your API key is not configured, you will be prompted to enter it. Alternatively, set the GEMINI_API_KEY environment variable.

Deploy Application

autokeren "deploy a simple shoe shop with HTML+CSS, using D1 for products"

The agent will automatically create the project, write the code, and deploy to Cloudflare Workers with D1 and R2 bindings.

Available Models

Alias	Model
kimi-code	Moonshot Kimi K2.7-Code (primary)
kimi-2.6	Moonshot Kimi K2.6
glm-5.2	Zai GLM 5.2 (secondary)
glm-flash	Zai GLM Flash
llama-4-scout	Meta Llama 4 Scout
gemma-4	Google Gemma 4
nemotron	NVIDIA Nemotron

Additional paths:

Alias	Model
gemini-3.5-flash	Google Gemini 3.5 Flash via AI Studio (--aistudio)
gemini-3.5-pro	Google Gemini 3.5 Pro via AI Studio (--aistudio)

Select a model with -m <alias>. Default: kimi-code with fallback to glm-5.2.

Commands & Shortcuts

Use the following keyboard shortcuts and slash commands in TUI mode:

Keyboard Shortcuts (Hotkeys)

Key	Action	Description
F1	Help	Toggle help dialog listing commands and shortcuts
F2	Switch Model	Open an interactive modal to switch AI models
F3	Reset Session	Reset the conversation history and tool permissions
F4	Copy Response	Copy the last AI message response to the clipboard
F5	Compact	Compact/summarize the conversation history
F6	Switch Language	Open a modal to change the TUI language interface
F7	File Explorer	Toggle the file tree sidebar on the left panel
Ctrl+K	Kanban Board	Toggle the project Kanban board panel
Ctrl+D	Debate View	Toggle the multi-agent background debate view
Ctrl+C	Cancel / Stop	Stop the active AI generation or running shell tool
Ctrl+Q	Force Quit	Force quit the autokeren CLI application

Slash Commands

Enter slash commands directly into the chat input box (Tab autocomplete is supported):

Command	Description
/help	Display help guidelines
/q or /quit	Exit the CLI session
/model [name]	Switch AI model (opens pop-up if name is omitted)
/lang [code]	Switch TUI language (opens pop-up if code is omitted, e.g. /lang id)
/export [name]	Export conversation to a Markdown file
/copy [last|N]	Copy a specific message to the clipboard
/mcp	Open the interactive Model Context Protocol (MCP) server manager
/mcp list	List active MCP servers
/mcp show <name>	Show tools from a specific MCP server
/mcp add <name> <cmd>	Add and start a new MCP server
/config	View active settings configurations
/config git-commit on|off	Toggle automatic git commits
/config cf-verify on|off	Toggle auto visual verify after deployment
/local [url]	Set/view local LLM endpoint (Ollama)
/approval on|off|ask	Set tool execution approval mode
/project <subcommand>	Multi-agent project management command
/compact	Compact conversation history
/reset	Reset the active session
/memory	View stored cross-session memory for the project
/permissions	View currently granted tool execution permissions
/save [name]	Save the current session state
/resume <name|id>	Resume a saved session
/sessions	List all saved sessions
/rewind [N]	Undo N tool calls and restore codebase to a checkpoint
/rewind list	List all available checkpoints
/genome	View codebase structural genome
/genome rescan	Rescan codebase architecture genome
/genome check	Scan for duplicate functions
/loop status	View loop breaker error history
/loop reset	Reset loop breaker statistics
/loop break	Break loop manually (swaps active model)
/review [staged]	Run cross-model code review
/security [file]	Run security audit scanner on a file
/spec <request>	Start requirements gathering interview
/spec answer <text>	Submit answer to interview question
/spec generate	Generate plan.md and technical-plan.md
/spec show	View implementation plan
/spec progress	Track implementation progress
/ghost <task>	Launch a background ghost agent
/ghost list	List all running background ghost agents
/ghost show <id>	View logs of a specific background ghost agent
/ghost kill <id|all>	Terminate background ghost agents
/research <query>	Search Reddit, Hacker News, and Google Web
/deploy <desc>	Create project and deploy directly to Cloudflare Pages/Workers

Stored Configurations

Config file is stored at ~/.config/autokeren/config.yaml.

auth:
  mode: "platform"       # "platform" (default), "direct", or "aistudio"
  api_key: ""            # API key from developers.autokeren.com
  gemini_api_key: ""     # Google AI Studio API key (only for "aistudio" mode)

cloudflare:
  primary_model: "kimi-code"
  secondary_model: "glm-5.2"
max_tokens: 16384
  temperature: 0.3
  timeout: 120.0

retry:
  max_retries: 5
  base_delay: 1.0
  max_delay: 60.0
  exponential_base: 2.0
  jitter: true
  circuit_failure_threshold: 5
  circuit_open_seconds: 30

autokeren:
  plan_mode: false
  max_iterations: 50
  shell_timeout: 180
  shell_allowlist: ["node", "npm", "pnpm", "npx", "git", "wrangler", "python3", "pytest"]
  project_root: "."
  context_window: 262144
  compact_tail_turns: 6
  auto_compact: false
  auto_compact_threshold: 0.8
  # Vibe coding features
  time_travel:
    enabled: true
    max_checkpoints: 50
    auto_checkpoint: true
  architecture_guardian:
    enabled: true
    block_duplicates: true
    scan_interval: 5
  loop_breaker:
    enabled: true
    max_repeats: 3
    auto_switch_model: true
  cross_model_review:
    enabled: true
    reviewer_model: "auto"
    auto_review: false
  vibe_security:
    enabled: true
    scan_on_write: true
    block_on_critical: true
  live_enforcement:
    enabled: true
    rules_file: ".ak-rules.yaml"
    block_on_violation: true
  spec_driven:
    enabled: true
    num_questions: 20
    auto_generate: true
  ghost_agent:
    enabled: true
    max_background: 3
    tmux_prefix: "ak-ghost"
  research:
    enabled: true
    sources: ["reddit", "hackernews", "web"]
    max_results: 10
    max_depth: 3
    summarize: true

mcp_servers:
  - name: filesystem
    enabled: true
    command: ["npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    env: {}

Environment Variables

Variable	Description
AUTOKEREN_API_KEY	Overrides configuration API key
GEMINI_API_KEY	Google AI Studio API key
AUTOKEREN_CONFIG_DIR	Custom config directory path (default ~/.config/autokeren)

Update

To upgrade to the latest version:

pipx upgrade autokeren

Go Runtime and pipx Distribution

autokeren runs interactive autokeren, prompt/task mode, --login, --init, --proof, sessions, tools, provider routing, and the Bubble Tea TUI through the native Go runtime by default.

pipx install autokeren remains the supported installer and updater. The Python package supplies the small cross-platform launcher and the matching compiled Go binary; it does not make Python the active agent runtime. pipx upgrade autokeren installs the next matching launcher and binary together.

Python is retained as an explicit compatibility path: use autokeren --engine python for legacy integrations such as custom roles, Telegram, or Google Antigravity. autokeren --engine auto starts Go first and may use the compatibility path only if the native provider path cannot start.

Contributing

Contributions are welcome! Please fork the repo, create a feature branch, and submit a PR.

git clone https://github.com/autokeren/autokeren.git
cd autokeren
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Before committing code, make sure ruff check ., mypy autokeren, and pytest all pass successfully.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Disclaimer

autokeren is an independent project and is not affiliated with, endorsed by, or sponsored by Cloudflare, Inc. "Cloudflare" and its associated product names are trademarks of Cloudflare, Inc. autokeren uses public Cloudflare APIs and workers infrastructure as a third-party client.