Maka records every agent turn as local execution data
Maka treats a coding-agent session as a saved sequence of model messages, tool calls, tool results, permission decisions, and termination events. The desktop and terminal interfaces read from that record, and later model calls receive a projection of it. Old tool output can be omitted from the next prompt without deleting the underlying evidence. That design is useful when a browser refresh, process crash, or context limit should not erase what the agent did.
The repository provides 3 surfaces: an Electron desktop app, a terminal interface with non-interactive runs, and an evaluation runner. All Maka work passes through Runtime Host, which owns sessions and tool execution. The evaluation layer builds task, repetition, and subject cells, then stores attempts, scores, usage, cost, duration, failures, and artifacts. This is more like an agent workstation than an editor extension.
Apple Silicon is supported, while 3 common targets are not
The README targets Apple Silicon macOS for the desktop release. Intel Macs and Linux are unsupported, and Windows x64 is an unsigned preview rather than a supported tier. Open issue #2142 tracks the Windows path, including signing, required CI checks, broader sandbox work, and Computer Use. A current Windows-specific issue, #3938, reports Glob failing when an approved project tree contains a nested NTFS junction.
Maka is also an Apache Incubator project. It has not produced an Apache-approved source release, and the README recommends no prebuilt download until that happens. Version 0.1.11, published August 18, 2026, is described as a convenience artifact produced during incubation. Organizations with a formal Apache release policy should treat the source build as evaluation material, not quietly relabel the existing installer.
The local record improves recovery but keeps secrets in plaintext
Runtime state lives in runtime.sqlite, with settings, artifacts, model connections, and a credential vault in the workspace directory. API keys are stored in a local plaintext JSON file restricted to the operating-system account. The renderer does not receive the keys, which limits one exposure path, but disk encryption and account access remain part of the security boundary.
Interrupted-turn resume is disabled by default. Enabling MAKA_RUNTIME_SAFE_BOUNDARY_RESUME=1 allows desktop, CLI, and startup resume behavior, and the README warns that resuming calls the model and consumes tokens. Older JSONL transcripts and Electron safeStorage credentials are not imported, so an upgraded workspace can appear empty and require credentials to be entered again. That migration boundary needs a backup plan.
What happened when we ran it
Our sandbox installed commit 2e8cd03 in 38 seconds, pulling 849 npm packages and occupying 1,420 MB. The monorepo build completed successfully in 23 seconds. Npm audit reported 0 known vulnerabilities across the installed dependency tree. The source checkout itself contained 2,774 files, about 853,049 lines, and 43.2 MB before installation.
The test command ran for 801 seconds and exited with code 1. Node's test report counted 1,092 passed and 3 failed out of 1,095. The supplied log tail says test:dist failed in @maka/runtime-host, then lists runtime, eval, desktop, storage, and runtime-host as failed workspaces. It does not include the 3 failed assertions, so we cannot say which behavior broke or why.
That distinction matters because the log also shows a very large passing set. The test run was close to green, yet a release gate that returns exit 1 is still a failure. Our Debian container had 3 CPUs, 8 GB of RAM, Node 22, no secrets, and no elevated privileges. It did not launch Electron or exercise macOS and Windows sandbox enforcement.
Built-in file tools sit behind explicit approval boundaries
Maka includes Read, Write, Edit, Bash, Glob, and Grep. Tools that leave the sandbox require approval, runs can be aborted, and failures receive classifications. Local memory and web search work when configured. Computer Use and catalog skills are optional rather than enabled by default, while the 0.1.11 release includes the Computer Use skill without its executor.
The graph mode can split implementation work across isolated Git worktrees. It requires the source project to be clean before it starts, which protects the worktree machinery but conflicts with the ordinary state of many development folders. Maka also supports MCP-related runtime behavior, and release 0.1.11 mentions hardened rediscovery. Issue #3939 reports a name collision between Maka's deferred tool_search connector and OpenAI Responses' native tool of the same name.
August activity is fast, and the contracts are still moving
GitHub recorded the last push on August 26, 2026, with 3,563 stars and 328 combined issues and pull requests. Release 0.1.11 says it contains 375 merged pull requests from 24 contributors. That is intense current activity. The README also warns that data formats, CLI commands, and experimental capabilities may change, so the same velocity raises upgrade work for early users.
The documentation is unusually candid about unsupported platforms, plaintext credentials, resume costs, data migration, and Apache release status. Our failed 801-second test command keeps this out of the dependable-default category, even with 1,092 cases passing. Maka is a good research and pilot choice for an Apple Silicon developer who values recoverable execution history. It is early infrastructure, and its own docs say so.