Your work. Your agent.

Maka keeps the task, the tools, and the evidence on one inspectable working plane.

Your work.Your agent.

A local-first Agent workspace built for real work—not another disposable chat window.

THE EXECUTION FIELD

01: Conversation
02: Tool Call
03: Artifact
04: Recovery

Localon your machine
Toolpermission-bound
Artifactready to inspect
Recoverprogress preserved

The work stays visible.

Maka connects the request, the Tool Calls, and the generated files in one workspace. You can inspect what happened and review the output beside the conversation that produced it.

Maka workspace showing a Chinese task conversation beside generated Markdown, diff, and HTML files in the Artifact pane. — Artifact pane / light theme / real visual-test fixture

Conversation: The task remains the center of the workspace.
Tool Call: Execution and permission state stay attached to the turn.
Artifact: Generated files appear as real outputs, not chat decoration.

The log is the runtime.

Messages, Tool Calls, Tool Results, and termination facts enter an event log. Sessions, UI, model context, and recovery are projections over that recorded execution.

01Request
The user's task enters a Turn.
02AgentRun
Model and tools advance the work.
03Event Log
Execution facts are appended.
04Projections
Context and UI read the same record.
05Recovery
Interrupted work can be understood.

Desktop / TUI / Headless
          ↓
SessionManager → AgentRun → Model + Tool Runtime
          ↓
Runtime Event Log → Context / Session / UI projections
          ↓
Task Event Log → TaskRun → Self-check / AHE evidence

Context is not history.

Tool Result pruning and model compaction can change what the next inference sees without treating recorded execution as disposable context.

01 / LOCAL-FIRST

Your records stay on your machine by default.

Choose a cloud API, compatible gateway, or local model connection. Maka does not bundle a shared model account.

02 / DURABLE TASKS

A task may outlive a Turn.

TaskRun and its event log make long-running work interruptible, inspectable, and exportable.

03 / PERMISSIONS

Tools operate behind an explicit boundary.

Schema validation, dynamic availability, watchdogs, abort, and error classification keep execution legible.

Inspect the complete architecture

Bring the connection you trust.

First run makes the boundary explicit: add an API, local-model, or supported account connection, test it, then choose the default you want Maka to use.

Maka first-run screen in Chinese listing provider connections including Claude, OpenAI, GLM, Kimi, DeepSeek, and Ollama. — First run / provider selection / real visual-test fixture

ConfiguredConnection details exist.

Send-readyThe runtime can actually use it.

ExperimentalPreview paths stay labeled as such.

Failure stays part of the record.

Missing files, unsupported previews, interrupted tools, and provider errors do not disappear behind a success animation. The interface names the boundary and keeps recovery possible.

Read the security boundary

Maka Artifact pane showing a deleted file and a clear Chinese error message explaining that the generated file can no longer be read. — Artifact error state / explicit boundary / real visual-test fixture

One runtime. Three ways in.

01
Desktop

Daily interaction, generated files, provider setup, permissions, search, branching, and recovery.
Electron + React
02
TUI / CLI

Work from the current project directory or run one non-interactive Turn with the same connections.
maka / maka run
03
Headless

Advance durable tasks with budgets, continuation, evidence export, and failed-task retry.
maka eval

Built in the open.
Ready to inspect.

Maka is under active development and currently targets people running from source or contributing to the project. Start with the code, the architecture, and the boundaries as they exist today.

Open GitHub Run from source

github.com / maka-agent / maka-agent