Open-source AI software engineer that runs in your terminal. Tell it what you need — it plans the work, spins up to 25+ parallel agents, controls the desktop with 28 tools (screen reading, mouse, keyboard, file system, process inspection), switches from Groq to an AMD MI300X GPU if one goes down, and keeps going until the task is done.
What changed from the old single-loop version to the current one.
| Capability | Swades v2.0 (Previous) | Swades v4.0 (Current) |
|---|---|---|
| Agent Scaling | Single ReAct agent loop only | Adaptive 4-tier scaling from 1 solo agent to 25+ subagents |
| Orchestration | Linear step execution | 9-Phase Loop (Analyse, Plan, Approve, Assign, Monitor, Debate, Merge, Verify, Summarize) |
| File Editing | Rigid exact-string matching | 4-Tier Fuzzy Patch Engine with boundary-drift & auto-creation |
| Edit Protocol | JSON tool calling only | Dual Ingestion: JSON tool calls AND plain-text Aider Search/Replace diffs |
| Merge Safety | Overlapping edits caused collisions | DAG dependency mapping with Kahn's topological sort & atomic file locks |
| Quality Control | Manual or basic syntax check | Automated quality gates (linters, typechecks, unit tests) with instant rollback |
| Memory | Ephemeral per-run memory | 4-Layer Persistent Memory (Project, Preferences, History, Telemetry) |
| Provider Support | OpenRouter only | Native Groq (300+ tok/s), OpenRouter, OpenAI, & 100% offline Ollama |
| Desktop Automation | Text CLI only | 28 tools: reads the screen (OCR), finds UI elements (AT-SPI2), types, clicks, scrolls, drags — watched via a read-only desktop stream |
| Inference Failover | Single provider, no fallback | Runs on Groq by default. If Groq is down, switches automatically to AMD MI300X GPU (192GB) with no interruption |
| Process Inspection | None | Reads /proc to list running processes, check open network connections, and get CPU/RAM/disk/GPU usage |
| Missing Dependencies | Manual — user has to fix | Runs apt install on its own when a package is missing, then picks up where it left off |
How it handles parallel agents, file edits, safety, memory, and desktop control.
Swades looks at the task and decides how many agents to use. A small fix gets one agent. A big refactor or security audit gets up to 25+ agents running at the same time, each working in its own isolated copy of the repo.
You decide how much the agent can do on its own. The desktop stream is read-only — the agent can only act through its tools, and every action is logged.
The agent runs without asking. Good for throwaway environments, fresh repos, or generating tests where you don't mind it doing whatever it needs.
Normal edits go through automatically. The agent pauses and asks before anything risky — installing packages, deleting files, or changing a database schema.
You approve every diff and shell command before it runs. Slower, but nothing happens without you seeing it first.
Loading...
Zero build steps. Zero runtime config choices. Run in your project directory:
npm install -g @xerv/swades-agent
NPM GLOBAL