“Do not make me configure this twice.”
ma init gives DeepSeek and LM Studio the same guided path: discover models, store credentials, create profiles, start working.
v0.2.0-alpha · Agora memory + reliable local-agent loops
MA
Your local coding agent should finish work, show evidence, and remember only when it is real.
MA makes local and remote models practical for real repositories: guided setup, hardened tool loops, managed dev servers, explicit context budgets, verified browser delivery, and native Agora MemoryPatch state.
95.3%
L2 pass rate proving local Qwen productivity
Agora
MCP stdio runtime with verified MemoryPatch state
Evidence first
tests and browser interaction before delivery claims
Managed loops
process lifecycle, byte budgets, and failure summaries
The Product Is The TUI
Model, tools, task state, context, and shortcuts are visible while the agent works.
First Run Matters
LM Studio, DeepSeek, or Agora: one guided path, no hand-written JSON.
The first screen is deliberately boring in the best way: arrow-key provider choice, sensible defaults, credential names, Keychain storage for remote keys, then model discovery.
What Users Actually Want
Less friction, lower cost, more usable local intelligence.
“Do not burn tokens for every loop.”
Local Qwen through LM Studio becomes the default place for repeated repo work, while DeepSeek stays one command away.
“Make my small model less dumb.”
Sampling passthrough, tool-call recovery, multimodal payload compatibility, and message integrity tests target Qwen/LM Studio behavior directly.
“Let it work for a long time.”
MA detects context windows, shows usage, compresses output, and is built for long agent loops rather than short chat demos.
“Do not leak my API key.”
Remote keys live in macOS Keychain, with config files storing references instead of plaintext secrets.
“Show me what the agent is doing.”
The TUI surfaces model, endpoint, tools, task progress, thinking state, session controls, and context budget while work is happening.
Benchmark As Evidence
The claim is not “we support local models.” The claim is “local models can deliver work you can verify.”
The benchmark is proof that MA's small-model engineering matters: local Qwen3-30B, LM Studio, tools, long multi-turn repo tasks, 70-task alpha gate. It is not a generic leaderboard. It is evidence for the product promise.
Read benchmark detailsRuntimeLM Studio
ModelQwen3-30B local
Tasks70
L0100%
L198.7%
L295.3%
Install
Download, extract, run.
macOS / Linux
tar -xzf ma-*.tar.gz
cd ma-*
./ma init
./maWindows
Expand-Archive ma-*.zip
cd ma-*
.\ma.cmd init
.\ma.cmd