Tools

Tools are how the model takes actions on your machine: read a file, run a shell command, search the web, query a database. Kenaz ships built-ins and connects to any Model Context Protocol (MCP) server you point it at.

Built-in tools

Built-in tools are shipped with the harness binary. The shell and network tools have an explicit on/off switch in the Tools view and are off by default. The filesystem tools don't have a separate switch — they're governed by the permission prompt at call time. kenaz__save_artifact and kenaz__update_artifact are on by default.

Tool	What it does	Gated by	Default
kenaz__read_file	Read a file by path	Filesystem-read permission	Permission-gated
kenaz__list_dir	List a directory's entries	Filesystem-read permission	Permission-gated
kenaz__glob	Glob-match files across a directory tree	Filesystem-read permission	Permission-gated
kenaz__grep	In-process regex search across files	Filesystem-read permission	Permission-gated
kenaz__write_file	Atomically create or overwrite a file	Filesystem-write permission	Permission-gated
kenaz__edit_file	Replace a substring in an existing file (requires a prior read_file call in the session)	Filesystem-write permission	Permission-gated
kenaz__bash	Run a shell command	Tools-view switch + Shell permission	Off
kenaz__web_search	Web search via your configured provider	Tools-view switch + Network permission	Off
kenaz__web_fetch	Fetch arbitrary HTTP URLs	Tools-view switch + Network permission	Off
kenaz__save_artifact	Persist a generated artifact to the Artifacts tab	(no permission needed)	On
kenaz__update_artifact	Revise an existing artifact in place	(no permission needed)	On

Each tool call is gated by permissions — see that page for how to pre-grant categories or paths so the model isn't pinging you constantly.

MCP — Model Context Protocol

MCP is the open spec for how AI clients (like Kenaz) connect to external tool/data servers. There are hundreds of MCP servers — official ones from major SaaS vendors, plus a long tail of community servers.

For step-by-step install instructions, examples (Filesystem, GitHub, Postgres, Slack, Docker-based, custom Python), JSON-config import, per-project configs, and how to write your own server, see Installing MCP servers.

The official catalog: github.com/modelcontextprotocol/servers.

Connecting an MCP server

Tools view → Add MCP server.

Three transports are supported:

• stdio — Kenaz spawns the server as a child process and talks over stdin/stdout. The most common pattern; fits anything you can run with npx or uvx.
• HTTP — point at a URL the server is listening on. Use for servers running as system services or in your team's infrastructure.
• SSE — server-sent events over HTTP. Used by some hosted MCP services.

For stdio servers, the command is what you'd run in a terminal:

# Filesystem (sandboxed to a path)
npx -y @modelcontextprotocol/server-filesystem /Users/you/code

# GitHub
npx -y @modelcontextprotocol/server-github

# Slack
npx -y @modelcontextprotocol/server-slack

You can usually copy these straight from the server's README.

Credentials for MCP servers

Some MCP servers need API keys (GitHub, Slack, Google Drive, …). Kenaz handles this two ways:

• Env vars on spawn. Type GITHUB_TOKEN=ghp_… in the env section of the server config; Kenaz writes it to the keychain and re-injects on spawn. Plaintext only ever exists in memory while the child process is alive.
• OAuth flow. Some servers (Notion, Slack, Google Drive) use the harness's credential resolver to mint per-spawn tokens via OAuth. Kenaz drives the OAuth flow once; subsequent spawns get fresh tokens automatically.

Tool discovery

When you connect an MCP server, Kenaz pulls its tool list and shows it in the Tools view. Each tool can be:

• Enabled by default — the model sees it on every session.
• Available — the model only sees it if you turn it on in the session.
• Disabled — never exposed.

The default for newly-discovered tools is Available (not Enabled) — so the model can't accidentally call a Slack-message-sending tool on its first turn. Promote to Enabled once you've vetted the tool.

Per-session tool selection

In the chat header → Tools dropdown — toggle individual tools on/off for that session. Useful when you want a focused agent (e.g. "filesystem only, no shell") without permanently disabling tools globally.

Recommended starter MCPs

For most users:

• Filesystem (@modelcontextprotocol/server-filesystem) — scoped reads/writes within a path. Better than the built-in for project work because you can restrict the root and control paths declaratively.
• GitHub (@modelcontextprotocol/server-github) — issues, PRs, commits.
• Web fetch (@modelcontextprotocol/server-fetch) — HTTP fetch for arbitrary URLs.

For developers:

• SQLite / Postgres — query a local DB.
• Memory (@modelcontextprotocol/server-memory) — persistent fact storage across sessions, useful for long-running agents.

For knowledge workers:

• Notion / Slack / Google Drive — read your documents, summarize threads.

Useful links

• Model Context Protocol spec
• Official server catalog
• Community server list
• Building your own MCP server — Python, TypeScript, Go, Rust SDKs are official