maka

Product Introduction

1. Definition: Apache Maka (Incubating) is a local-first, open-source AI agent workspace and runtime environment. Technically, it is an event-sourced system designed for the development, execution, and audit of autonomous AI agents.
2. Core Value Proposition: Maka exists to solve the critical problem of opaque and unaccountable AI agent execution. Its primary value is providing developers and teams with full auditability, transparency, and control over AI agents by immutably logging every model interaction, tool call, and decision.

Main Features

1. Immutable Runtime Event Log: The core architectural component. All agent activities—model messages, tool calls, tool results, permission decisions, and termination events—are recorded as an append-only, recoverable ledger. This log serves as the single source of truth, enabling complete session replay, audit trails, and debugging.
2. Local-First Architecture & Model Agnosticism: Sessions, settings, and execution records are stored on the user's machine by default. Users retain full control over their data and can connect to any model provider (cloud API, local model, or compatible gateway) without vendor lock-in, enhancing data privacy and flexibility.
3. Controlled Tool Execution with Permission Engine: Agents can perform real work using local tools like Read, Write, Edit, Bash, Glob, and Grep. A dynamic permission policy engine validates tool schemas, manages availability, and enforces security boundaries for dangerous operations like file writes and shell commands.
4. Multi-Surface Interface (Desktop, TUI, CLI): Provides tailored entry points for different workflows. The Electron-based Desktop app offers a rich UI for daily interaction and artifact workflows. The Terminal User Interface (TUI) and non-interactive CLI (maka run) allow integration into development scripts and CI/CD pipelines, sharing the same runtime host.
5. Structured Evaluation Framework: A declarative system for running reproducible benchmark experiments. It supports multi-arm experiments (task × repetition × subject), immutable per-cell attempts, and a results kernel for scores, normalized token usage, cost, duration, and artifacts, enabling rigorous AI agent performance testing and comparison.

Problems Solved

1. Pain Point: Lack of AI Agent Transparency and Auditability. Traditional agent frameworks operate as black boxes, making it impossible to verify why an agent made a specific decision or took a particular action, which is a major barrier to trust and deployment in production environments.
2. Target Audience: AI Engineers and Developer Teams building production-grade autonomous agents; Security and Compliance Officers who need verifiable logs of AI actions; Researchers requiring reproducible experiments and detailed traces of agent reasoning for analysis.
3. Use Cases: Debugging and Improving Agent Logic by replaying failed sessions from the event log; Compliance and Security Audits by providing an immutable record of all AI-driven actions and data accesses; Cost and Performance Optimization by analyzing detailed usage metrics per session and tool call; Comparative Evaluation of different AI models or agent prompts using the built-in evaluation framework.

Unique Advantages

1. Differentiation: Unlike cloud-hosted agent platforms (e.g., LangChain/Smith in cloud mode) or simpler orchestration frameworks, Maka's local-first, event-sourced core prioritizes user sovereignty, data privacy, and forensic-level observability. It treats the execution log as the runtime itself, not just a side effect.
2. Key Innovation: The separation of the durable Runtime Event Log from contextual projections. This allows features like "Tool Result pruning" and "LLM Compaction" to manage what the agent sees next without discarding recorded evidence. This architectural choice ensures auditability is not sacrificed for context window management.

Frequently Asked Questions (FAQ)

1. What is Apache Maka used for? Apache Maka is used for developing, operating, and auditing transparent AI agents. It is essential for developers who need a verifiable history of their AI agent's actions, reasoning, and tool usage, particularly in scenarios requiring accountability and reproducibility.
2. Is Maka a hosted service or a local tool? Maka is a local-first workspace. By default, all data—including session history, execution logs, and agent artifacts—resides on your local machine. You connect it to your own choice of AI model APIs or local LLMs, maintaining full control over your data and workflows.
3. How does Maka ensure the security of tool execution? Maka employs a runtime permission engine that acts as a security boundary. It validates tool schemas, dynamically controls tool availability based on policy, and requires explicit user consent for dangerous operations like shell commands or file writes, preventing unauthorized agent actions.
4. Can I use Maka to compare different AI models? Yes, Maka includes a dedicated evaluation framework specifically designed for running reproducible, multi-arm benchmark experiments. You can declaratively define tasks and test different AI models (subjects) to compare their performance, cost, and effectiveness in a controlled manner.
5. What happens if an agent execution is interrupted? With the MAKA_RUNTIME_SAFE_BOUNDARY_RESUME=1 flag enabled, Maka's Runtime Host supports continuation and recovery. The Desktop app, CLI, and TUI can resume interrupted turns by replaying the immutable event log, allowing agents to recover from crashes or stops without losing context.