a coding agent
A single python script that edits code, runs terminal commands, reads images, and tests its own work. Zero extra packages for core, with 6 modular plugins.
How it works
The agent checks your workspace, decides what needs to be changed, and updates files step by step.
Look around
Checks folder layout, reads source files and images (PNG/JPEG), searches with regex, and outlines syntax.
Check status & safety
Compares real files (+, ~, -) and git changes. Guards network fetches against SSRF and scrubs sensitive env tokens.
Plan & delegate
Figures out exact code edits. Can spawn isolated child sub-agents with all tools or a restricted tool list.
Edit, test & verify
Patches lines cleanly, runs terminal commands or tests, and an independent critic verifies the result.
Core Tools & Plugins
13 built-in tools in the core script, plus 6 modular plugins when extra capabilities are needed.
- map_repo & find_files — File tree, sizes & glob search
- grep & analyse_file — Regex search & AST outline
- read_file — Reads code files and decodes images (PNG, JPG)
- create_file & patch_file — Creates files & edits lines cleanly
- run & delete_file — Shell execution with timeouts & safe deletion
- web_search & web_fetch — Web search & page fetching with image support
- Groq & OpenRouter — Fast LPU speed or wide model selection
- spawn(task, tools) — Sub-agents with full or filtered tool access
- plan & state — Step planning and real workspace file tracking
- ask — Pauses to ask the user clarifying questions
- critic verdict — Independent reviewer verifies proof before done
- cost ledger — Per-run tokens, costs, and monthly spending budget
- hardware — ESP32 pan servo camera head (hw_look, hw_pan)
- cua — Desktop automation via xdotool & screenshots
- browser — Headed or headless web browser via Playwright
- mail — IMAP/SMTP client with draft approval outbox queue
- notes — Persistent scratchpad and inter-run memory
- rest — Authenticated HTTP/REST caller with host allow-list
- SSRF protection — Blocks loopback (127.0.0.1, ::1) and private IPs
- env scrubbing — Cleans sensitive tokens before subprocess runs
- path guard — Prevents edits to the agent's home directory
- stop switch — Immediate pause via python agent.py --stop
- loop breaker — Stops automatically if stuck repeating edits
- offline selftest — Tests tools, critic, plugins, and mock APIs
Why we kept it simple
Instead of complex frameworks, Swades is built as one straightforward python file.
| Feature | Typical agent setups | Swades |
|---|---|---|
| Setup | Many python packages to install | 1 single script (zero pip install required for core) |
| Providers | Single provider lock-in or heavy SDKs | Dual providers: Groq (LPU speed) & OpenRouter (broad models) |
| Extensions | Editing core engine or complex class inheritance | Drop python files into plugins/ without modifying core script |
| Images & Vision | Separate microservices or extra vision dependencies | Automatic PNG/JPEG decoding in read_file and web_fetch |
| Verification | Assumes model output is correct | Independent critic re-checks evidence before marking done |
| Offline testing | Needs live API and network | Built-in --selftest tests all tools offline with mock servers |
| Safety & budget | Unbounded execution and open internal network calls | SSRF blocking, env scrubbing, monthly cost ledger & STOP switch |
Recent changes
Loading...
Run Swades
Run directly in your terminal or Google Colab with python:
git clone https://github.com/Xerv-Org/Swades.git && cd Swades && python agent.py --selftest