Web Forge

Web Forge is IsonForge in your browser. Same engine as the CLI, different surface.

What you get

Virtual workspace in your browser. Drag-drop files in; they live in IndexedDB until you download them.
Isolated bash sandbox per session. Real Linux environment with python, node, git, pytest, etc. Network on. Workspace files mirrored in.
Same agentic loop as the CLI: tool calls, diffs, plan mode, sub-agents, hooks (via gateway settings), Skills.
Image attachment via drag-drop or paste.
Auto-checkpoint at session idle. Restore picks up where you left off.

Why use Web Forge

No install. Browser only. Good for casual use or when CLI install is overkill.
Visual diffs. Inline unified-diff preview before applying edits.
Share-friendly. Link a session for review.
Mobile-OK. Slower input, but viable on a tablet.

When to use CLI instead

You're working in a real project directory and want IsonForge to edit your real files.
You need MCP servers (CLI-only currently).
You're running automated / scheduled tasks (-p, --bg).
You're using gh CLI for PRs.

CLI and Web are siblings - they don't share state. A CLI session in ~/projects/myrepo and a Web Forge session with the same files are independent.

Workspace tools

In Web mode, the agent has a different toolset focused on the virtual workspace:

Tool	What
`workspace_list`	Enumerate files in the workspace
`workspace_read`	Read a file
`workspace_write`	Create or overwrite a file
`workspace_edit`	Surgical string-replace edit
`bash`	Run in the sandbox container
`web_search`, `web_fetch`	Same as CLI
`todo_write`	Plan tracking, shown in sidebar
`ask_user`	Clarifying questions with button UI

The agent uses workspace_* to modify files (changes appear in your browser UI) and bash to run tests / lint / build (output streams to the REPL). Bash modifications inside the sandbox don't persist back to the workspace - use workspace_write to update browser-visible files.

See Virtual workspace and Bash sandbox for details.

Plan mode

Toggle plan mode from the UI or via /plan. Same enforcement as CLI: workspace_write, workspace_edit, and bash are stripped from the tool list. Agent can only read + research + propose.

File limits

~1 MB per file (workspace).
~50 MB total workspace.
8 MB per image attachment.

Files larger than these are rejected at drag-drop time.

Session persistence

Web Forge sessions auto-snapshot at 30-minute idle. Reopen the URL and you're back where you left off, including the sandbox container state (it's docker commit-ed as a snapshot image and restored in 1-2 seconds).

Snapshot count caps at 20 - oldest gets evicted. Snapshots are roughly 280 MB each.

Drag-drop

Drag any combination of:

Source files (.py, .js, .ts, .md, etc.)
Images (.png, .jpg) - vision-described by the agent.
Zip archives - extracted into the workspace.

Files appear in the workspace sidebar. The agent sees them via workspace_list and workspace_read.

Download

When done, hit "Download workspace" to get a zip of every file. Includes anything the agent created via workspace_write.

Slash commands

Web Forge supports a subset of CLI slash commands:

/clear
/compact
/undo
/init
/plan
/web <query>
/help
/checkpoint, /rollback, /checkpoints
/cost

It doesn't have CLI-specific ones (/sessions, /resume, /loop, /agents subcommand, etc.) because the surface is different.

Auto-compact

When context fills past ~78% of the model's window, Web Forge auto-compacts older turns. Reasoning + recent tool results stay; older raw chat history gets summarized.

Caveats

Browser memory. IndexedDB workspace = browser memory. Huge files or many sessions = slower browser.
No native FS. Web Forge can't touch files outside the workspace. Use CLI for that.
Sandbox is single-tenant. Don't share session URLs with strangers - they get container access.