Core concepts
The handful of ideas that show up in every other page. Read this once and the rest of the docs makes sense.
Workspace
A workspace is one project — a directory on the KrowForge server with its own files, terminal sessions, agent history, secrets, and settings. You can have many workspaces; each one is fully isolated.
Workspaces live under workspaces/<slug>/ on the server. Everything you do is scoped to the active workspace.
The agent
The agent is the LLM-driven assistant in the right panel. It has tools (read files, write files, run shell, search code, drive a browser) and follows a reasoning loop. When the agent wants to change a file, it doesn't write directly — it stages a diff, you review it, you accept or reject.
Hunks, snapshots, receipts
Three concepts that together make every change reviewable, reversible, and auditable:
- Hunk — a contiguous chunk of a unified diff. The smallest unit you accept or reject in the Changes panel.
- Snapshot — a copy of every touched file taken before an apply. Lives in a per-workspace ring buffer. The unit of rollback.
- Receipt — an immutable JSON record of every mutation, hash-chained to the previous receipt. The unit of audit.
Every apply produces a snapshot + a receipt. Every rollback produces another receipt.
Autonomy modes
How much the agent is allowed to do without asking you:
| Mode | Behavior |
|---|---|
| Manual | Agent plans only. Never proposes diffs without explicit ask. |
| Hybrid (default) | Agent proposes diffs. You always review before apply. |
| Autonomous | Agent applies safe diffs automatically (within budget caps + guards). |
See Autonomy modes.
Budget
Every agent run consumes tokens which cost money. KrowForge tracks cost per run and per day, with hard caps you set in settings. The composer shows a live cost estimate before you send.
See Budget & cost.
Guards
Three guardrails that fire on every write:
- Write guard — rejects paths outside the workspace, oversized files, binary files, dangerous patterns.
- Hallucination guard — flags references to functions/classes that don't exist in the symbol index.
- Egress guard — strips API keys, private keys, and out-of-project paths from agent frames before they reach your browser.
See Guards.
MCP
Model Context Protocol. The standard the agent speaks to its tools. Means you can plug in third-party MCP servers and the agent can use them too.
See MCP tools.