English | 中文
DeepSeek in minutes. Local small models turned into real productivity.
MA is a terminal coding agent built around two practical promises: remote setup should be brainless, and local small models should become useful production tools. DeepSeek config is interactive and direct. LM Studio/Qwen gets long-context handling, tool hardening, model switching, and benchmark-driven fixes so small models can do real repo work.
v0.1.2-alpha supports LM Studio local models and DeepSeek official API today. More OpenAI-compatible providers are next.
Website: https://zhuqingyv.github.io/my-agent/
Release: https://github.com/zhuqingyv/my-agent/releases/tag/v0.1.2-alpha
- Local small models become productive: MA's alpha gate runs 70 L0-L2 tasks through a local Qwen3-30B model via LM Studio.
- DeepSeek is the zero-friction fallback:
ma initgives LM Studio and DeepSeek the same arrow-key setup flow, stores remote keys safely, and leaves you with a working profile instead of a config chore. - Near-infinite working room: MA auto-detects context windows, tracks usage, compresses output, and is designed for long local-agent loops.
- Small-model hardening is the product: Qwen/LM Studio-specific sampling, image payload compatibility, tool-call recovery, and prompt/message integrity are treated as release gates.
- Agent tools are built in: shell, file read/write, structured edits, grep, and web are available immediately after init.
Most terminal AI tools assume the hosted model is the product. MA assumes the workflow is the product: configure DeepSeek without thinking, then make local Qwen useful enough to keep running real tasks without worrying about token cost.
That means the product priorities are different:
- DeepSeek setup that writes a usable profile in one pass
- local model profiles instead of one global model string
- benchmark gates for local small-model productivity instead of vibe-only demos
- Keychain-backed secrets instead of plaintext API keys
- repo-local instructions, skills, and tool loops tuned for small models
MA uses benchmark data as product evidence: local Qwen3-30B through LM Studio passes the alpha L0-L2 release gate.
| Model | Runtime | Tasks | L0 | L1 | L2 |
|---|---|---|---|---|---|
| Qwen3-30B local | LM Studio | 70 | 100% | 98.7% | 95.3% |
This benchmark is the proof point for the claim: local small models can become useful with enough agent-loop engineering. It covers connectivity, stable tool use, and multi-turn local project work. It is not a universal coding-agent leaderboard.
See docs/benchmark-results.md.
Download the release asset for your platform:
ma-*-macos-arm64.tar.gzma-*-linux-x64.tar.gzma-*-windows-x64.zip
macOS / Linux:
tar -xzf ma-*.tar.gz
cd ma-*
./ma init
./maWindows:
Expand-Archive ma-*.zip
cd ma-*
.\ma.cmd init
.\ma.cmdThe portable bundle includes Node.js and production dependencies. No global Node or npm install is required.
git clone https://github.com/zhuqingyv/my-agent.git
cd my-agent
npm install
npm run build
npm link
ma init
mama init
maDuring init:
- Choose model source: LM Studio local or DeepSeek official.
- Enter base URL if needed.
- Enter API key for remote providers.
- Pick a discovered model with arrow keys.
That means both first-run paths stay obvious:
LM Studio local -> Base URL -> credential name -> discovered local model
DeepSeek official -> Base URL -> credential name -> Keychain API key -> discovered DeepSeek model
Inside MA:
/ show slash command suggestions
/model switch model/profile with arrow keys
Tab complete selected command
Enter run selected slash command
ESC ESC switch session
User-facing slash commands:
| Command | Purpose |
|---|---|
/model |
Open the model/profile picker |
/help |
Show user-facing commands |
/clear |
Clear current conversation |
/exit |
Exit MA |
CLI commands:
ma # chat
ma chat --resume # resume latest session
ma chat --resume <id> # resume specific session
ma sessions # list sessions
ma profiles # list model profiles
ma profile use <profile> # set default profile
ma secrets list # list secure credentials
ma secrets view <id> # view masked key after system auth
ma secrets delete <id> # delete key after system auth
ma secrets repair <id> # repair macOS Keychain trusted access
ma init # interactive setup
ma versionMA separates credentials from model profiles.
Example model ids:
LMStudio-local/qwen/qwen3.6-27b
DeepSeek/deepseek-v4-flash
/model aggregates models from configured providers, prefixes them by credential/provider name, and remembers the last selected profile.
MA starts with built-in MCP servers:
exec: shell command execution with danger guardfs: file read/writefs-edit: structured file editsgrep: code/text searchweb: DuckDuckGo search and web fetch with curl fallback
Create .ma/skills/deploy.md:
---
name: deploy
description: Deploy this project
arguments:
- name: environment
description: Target environment
required: false
default: staging
---
Deploy this project to {{environment}}.
Run tests first, build, deploy, then verify.Use it:
/deploy environment=production
Skills appear in slash command suggestions unless they conflict with a built-in command.
Global config:
~/.my-agent/config.json
Project config:
./config.json
Project config overrides global config. AGENT.md files are loaded from the current directory upward, plus ~/.my-agent/AGENT.md.
MA can run shell commands and edit files. Use it in trusted workspaces.
Current safeguards:
- dangerous shell command confirmation
- macOS Keychain for remote API keys
- explicit
ma secrets view/deleteauthentication - session-local runtime secret loading for unattended agent work
Known alpha boundary: the current Keychain helper is good enough for local alpha use, but stricter process-level trust would require a signed helper/ACL design.
npm run dev
npm test
npm run build
npm run release:checkMIT