Agent + MCP tools

The KrowForge agent is a chat surface backed by a tool-using LLM. The tools follow the Model Context Protocol so the same agent can swap models without rewiring.

How the chat works

Open the Ask tab in the activity rail. Type a message. Send. The agent streams a response.

What's different from a plain chatbot:

  • Tools. The agent can read files, write files, run shell commands, search the codebase, and drive a browser.
  • Diff review. When the agent writes files, it doesn't write them directly — it stages a diff. You review hunks in the Changes panel and accept what you want.
  • Rollback. Every accepted apply creates a receipt. One click reverts.
  • Subagents. For multi-step work, the agent can spawn focused subagents that report back a single message. You'll see them in the run log.

What tools the agent has

You can ask the agent: *"List your tools."* Verbatim list as of this build:

ToolWhat it does
read_fileRead a file by path.
write_fileStage a file write (goes through diff review).
apply_patchStage a multi-hunk patch.
list_dirList a directory.
grepRegex search across the project.
search_codebaseSemantic + lexical hybrid search over the embed index.
krowforge.semantic_searchLegacy alias of search_codebase.
run_shellExecute a shell command in the PTY.
browser.smoke / marketing / flowDrive the headless browser (allowlist gated).
notebook.*Notebook cell CRUD + run.

How to ask well

The agent is good at:

  • Bounded edits — *"In services/auth.py, change the password hashing from bcrypt to argon2."*
  • Multi-file consistent renames — *"Rename validate_user to verify_user everywhere it's used."*
  • Search-then-summarize — *"Where is the database connection string read from?"*
  • Diagnose-then-propose — *"This test is failing; tell me why and propose a fix."*

The agent is less good at:

  • Vague design questions without a target file. Specify scope.
  • Anything requiring the network outside the browser-tool allowlist.
  • Long-running operations the agent has to wait on. Use the terminal yourself for those.

Cost & model routing

Each turn picks a model based on the task. Cheap router for retrieval; mid-tier for code edits; top-tier for hard reasoning. You can see which model handled which step in the run log.

If a response feels under-cooked, ask *"Try again with a stronger model"* and the agent will re-route.

MCP — the external tool protocol

The agent's tools are exposed via the Model Context Protocol. That means:

  • Other MCP-aware agents can call KrowForge tools (when permitted).
  • KrowForge can call third-party MCP servers as additional tools.

Configuration lives in the workspace settings panel under MCP servers.