Skip to main content

Using the Harness

Once you've completed the Quickstart you have a working setup. This page is about the daily rhythm — how sessions, tools, memory, and permissions fit together during real work.

Starting a session​

Click New session in the Sessions view (or ⌘N / Ctrl+N). A session is one conversation with the model — it persists across restarts, can be branched, and can hold file attachments and produced artifacts.

If you're working on a recurring project, create a Project first (Sessions view → Projects → +). Sessions inside a project inherit a shared default model and system prompt, and the filesystem tool scopes reads to the project root by default.

tip

Give sessions descriptive names. Kenaz auto-suggests one from the first few turns, but renaming it yourself makes search much faster later.

Working with the model​

Type in the input box and hit Send. While the response streams in you can:

  • Attach a file — drag-and-drop onto the input, or click the paperclip. Images and PDFs go inline; other files get summarized. Attachment support depends on the model.
  • Switch the model — open the model picker in the chat header to try a different model mid-conversation.
  • Use a slash command — type / to open the command picker. Built-ins like /model, /branch, and /clear are always there; any custom commands you've saved show up too.

When the model wants to take an action — read a file, run a shell command, call an MCP tool — a permission prompt pops up before anything runs. Approve the single call, approve all for the session, or deny. See Permissions for how to pre-grant categories you trust so you stop seeing the same prompt repeatedly.

Branching and exploring​

Hover any assistant message and click Branch from here to fork the conversation at that point. The branch starts with all prior turns intact; you can take it in a different direction without losing the original. Branches are cheap — use them when you're unsure which approach to take.

Tool use​

The Harness ships built-in tools for filesystem access, shell commands, web search, and web fetch. The shell and network tools are off by default — turn them on in the Tools view when you need them.

To connect external tools, open Tools → Add MCP server and point it at any MCP server. Once connected, the model can call that server's tools exactly like built-ins. Newly-discovered tools default to Available (not auto-enabled) — promote them to Enabled in the session header after you've verified what they do.

note

Every tool call — whether built-in or MCP — is recorded in the audit log with the full arguments, result, and permission decision.

Permissions that don't slow you down​

The default is "ask each time." After a few sessions you'll know which paths and categories you trust. A useful starter setup:

  • Filesystem read → Always allowed → Project root only. The model reads your code without prompting.
  • Filesystem write → Always allowed → Project root only. Same scope.
  • Shell → Ask each time. Leave this one alone until you know the exact command pattern you want to approve.
  • Network → Approved for session. Web search is usually fine.

Change these any time from Settings → Security → Permissions.

Memory and long context​

Long conversations compact automatically when they approach the model's context window. Kenaz summarizes the oldest stretch and shows a [Summary] card in the chat. If you want to tune when this happens, adjust the aggressiveness in Settings → Authoring → Compaction.

Facts worth keeping across sessions go in long-term memory. Right-click any line → Remember this, or open the Memory view and add a chunk manually. Scope it to the session, project, or globally — the model retrieves relevant chunks automatically on future turns.

For team-wide facts that should be in front of the model every turn (not just retrieved), look at Context packs.

Artifacts — files the model produces​

When the model produces a file (code, a CSV, a diagram, a document), Kenaz captures it as an artifact in the Artifacts tab. You can preview it, download it, or promote it to project scope so other sessions can reference it. The model can also revise artifacts in place across turns.

Finding things later​

⌘F / Ctrl+F searches across session titles, message bodies, attachment filenames, and artifact names. Jump to the matching message directly from the result.

Closing the loop: the audit log​

The Audit log view shows every event the Harness recorded — prompts, responses, tool calls, permission decisions. Use it to verify what happened, export a report, or satisfy a compliance requirement. See Audit log for browsing, verifying the hash chain, and configuring retention.

Next: Explore the Features reference for deep-dives on each capability, or jump to Providers if you want to add a second AI provider.