Why this matters
Long-running or tool-using AI agents need more than better prompts: they need a runtime that preserves evidence, handles partial tool effects, enforces permissions, and can resume interrupted work. Maka treats the append-only Runtime Event Log as the single source of truth so sessions, recoveries, and experiments are rebuildable and auditable rather than ephemeral.
What Sets It Apart
- Append-only execution records: every model message, Tool Call, Tool Result, permission decision, and termination event is persisted before projection into UI state, enabling inspection, replay, and verification of agent behavior.
- Local-first control plane: workspace data, credentials (on-disk vault), and runtime storage live on the user’s machine by default while model connections can be cloud APIs, local models, or gateways — giving users choice over where inferences run.
- One Runtime Host for all surfaces: Desktop (Electron + React), TUI/CLI, and a headless Eval runner share the same Agent Runtime, tools (Read, Write, Edit, Bash, Glob, Grep), permission engine, and AgentRun ledger for consistent behavior across interactive and benchmarked runs.
- Context management without losing history: active Tool Result pruning and LLM compaction reduce what is sent to models while preserving original execution records for recovery and audit.
Core Capabilities
- Durable execution: runtime.sqlite is the operational authority holding events, Agent Graph state, artifacts, usage and continuation metadata, enabling opt-in safe resume and recoverable long-horizon tasks.
- Tooling and controls: built-in local tools, tool schema validation, dynamic availability, watchdogs, aborts, and permission pauses help contain side effects and classify failures.
- Evaluation-first ergonomics: declarative multi-arm experiments expand into immutable task×repetition×subject cells; Maka subjects execute through Runtime Host for reproducible benchmarking and attributable cost/usage reporting.
- Developer workflow: desktop app for daily interaction, TUI/CLI for workspace-driven sessions and non-interactive runs, and a headless runner for durable tasks and evaluation subjects.
Who It's For & Trade-offs
Great fit if you need auditable, recoverable agent runs (e.g., automation that executes code or makes filesystem changes), reproducible agent-based experiments, or a local-first workspace where you control model connections. Look elsewhere if you require a fully polished cross-platform desktop release today (current public desktop targets Apple Silicon Macs first), if you need a managed hosted service out of the box, or if you cannot accept local plaintext credential storage under the OS account boundary without additional hardening.
Where It Fits
Maka sits between lightweight LLM clients and full orchestration platforms: it’s a runtime and workspace for agents that combines UI surfaces with durable execution semantics, intended for engineers and researchers who run tool-using agents, long-horizon automations, or reproducible agent evaluations.
