Repository navigation
Releases: yogeshwardev/MYCODE
Releases · yogeshwardev/MYCODE
Release list
SwarmCode 1.1.0
SwarmCode 1.1.0 adds plan mode, AI code review for CI, guardrails, hooks, web search and image input, and cuts token use on long tasks.
Download
| Platform | File |
|---|---|
| Windows 10/11 (x64) | SwarmCode-Setup-1.1.0.exe |
| macOS (Apple Silicon) | SwarmCode-1.1.0-arm64.dmg |
| Linux (x64) | SwarmCode-1.1.0.AppImage or swarmcode-desktop_1.1.0_amd64.deb |
CLI on any OS: npm i -g github:yogeshwardev/MYCODE
These builds aren't code-signed yet:
- Windows: if SmartScreen appears, click More info, then Run anyway.
- macOS: if it says the app is damaged or can't be opened, move it to Applications and run
xattr -cr /Applications/SwarmCode.apponce. - Linux AppImage:
chmod +x SwarmCode-1.1.0.AppImage, then run it.
New
- Plan mode (
swarm plan,--plan,/planin the app): the agent investigates read-only and proposes a plan; nothing changes until you approve it. swarm review(/reviewin the app): reviews uncommitted changes, a branch (--base main) or a commit (--commit), never edits, and exits with code 2 when it finds issues, so it can gate CI.- Permission rules (
rules.allow/rules.deny, e.g."bash:npm test*","write:src/**"). Deny rules come from any config; allow rules only from your global config, so a cloned repo can't pre-approve its own commands. - Destructive-command guard: force pushes,
git reset --hard, recursive deletes of root/home, disk formatting, dropping databases and piping downloads into a shell always need a person, even in auto mode, and the approval prompt says why. - Hooks:
hooks.afterEditruns commands (formatters, linters) on each file the agent writes; failures go back to the agent. - Web search (
web_search): Brave or Tavily with an API key, DuckDuckGo otherwise. - Images:
--imagein the CLI, paste or drop screenshots in the app. @filementions attach files to a request.- Checklist (
todo) for multi-step work, shown live in the terminal and the app. - Clarifying questions (
ask_user): the agent can ask one short question with suggested answers instead of guessing. swarm init --agentswrites an AGENTS.md for the project.
Better
- Much lower token use on long tasks: older tool output is trimmed from context in batches (recall still finds it), and
contextBudget(default 160k) caps the working context even on 1M-token models. One benchmark task went from 2.18M to 0.59M tokens. - Works better with any model: common tool and argument names from other agents (
file_path,old_str,query,print_tree, ...) are understood, no-op edits no longer count as errors, and an agent that stops without replying or mid-step is told to continue. - Strongly matching skills are included automatically; skill triggers cover more everyday requests.
- Browser tools: type by field label, and clicks by text prefer buttons and links.
- Smaller, faster desktop app: installer 125 MB to 101 MB, minified bundles, startup about 1 s.
Fixed
- Parallel workers, merge and fix agents, and best-of candidates now follow deny rules and the destructive-command guard (they approved everything before).
- A folder inside a git repo at your home directory is treated as its own project everywhere (
initwrote to~,cleantouched the home repo); paths are compared as real paths on Windows. - CI: Node 22+ is required (the PDF library needs it); release builds no longer fail when signing secrets are absent.
SHA-256
0a5c2af2e575602178e528301fd7c9a14a64055a41fd0884fcb15fc76a7473c1 SwarmCode-1.1.0-arm64-mac.zip
9d20ee6eadfbdfefc908f9f19d3d72b9b7f07264347fd85551f62d6d1cc0771d SwarmCode-1.1.0-arm64.dmg
05d21779632cd06fbaa540a7e4b4f41c5ae0025f41935567ec1d9e6d04766b8f SwarmCode-1.1.0.AppImage
bb707eb3dfcc500d0682d9ced1d667ba972b43a56e85ecabd940596e3c9cdbb4 swarmcode-desktop_1.1.0_amd64.deb
5e362f456615ed61d5411d1853baaab301da3736b75779d964f8b771338b75cf SwarmCode-Setup-1.1.0.exe
SwarmCode 1.0.0
SwarmCode 1.0.0, the first public release.
Download: SwarmCode-Setup-1.0.0.exe below (Windows 10/11, 64-bit).
CLI on any OS: npm i -g swarmcode
Highlights
- Parallel agents: plan, build in isolated git worktrees, merge, test, review, apply
- Hidden-test guarantee: acceptance tests written before the code, never shown to the builders
- Best of 3 across models, security gate, proof report for every change
- Works with Claude, GPT, Gemini, DeepSeek, OpenRouter, Ollama and any OpenAI-compatible model
- 35 built-in skills, chosen automatically or with
/ - Browser testing, PDF/Word reading and PDF generation
- GitHub and Google sign-in
Install note
This build is not code-signed yet. If Windows SmartScreen appears, click More info, then Run anyway.
SHA-256: 78b6fcc5683749b5fc94d956a1e01c7c7a86a5dc29a5467e5ca55b4cfffe28d0