Coding Agents

The desktop app has the Claude Code tab: a local AI workspace where an agent edits files, runs commands, and reports results back to Liza. Despite the name it now hosts several providers — Claude, Codex, the included Liza AI, and Perplexity.

Desktop only

Coding agents run only in the desktop app, since they operate on a local working directory. In the browser the Liza assistant is available, but no coding agent.

What an agent can do

• Read and edit files in a working directory
• Run commands (terminal/bash)
• Use the project context
• Start and stop local dev servers and read their logs (see below)
• Report back to Liza (via MCP): check off tasks and post a summary to the chat

Connect a provider

Settings → Personal → AI settings → General

There you pick the provider at the top (one sub-tab per provider) and connect it – either by login or with your own API key. You can switch providers later at any time via the pills in the Claude Code tab.

Liza (included)

No setup needed: runs on the AI operated by Liza and counts against the AI credit. Available on paid plans if the organization has allowed it.

Several models operated in Europe are available – GLM-5.3 by default, plus Flash variants for simple work and stronger models for demanding work. They differ considerably in price; see Models of the included AI.

Claude (Anthropic)

• "Connect with Claude" – connects the desktop app to your Claude account (token stored locally, never sent to Liza), or
• Manually: paste your own key/setup token (claude setup-token or sk-ant-…). Create a key at console.anthropic.com.

Claude appears in the Claude Code tab in two variants: "Claude Code" (coding agent with file access) and "Claude" (chat/research without file access).

OpenAI / Codex

• "Connect with ChatGPT" – browser login (recommended if you have a ChatGPT subscription, no API key needed), or
• paste an API key from the OpenAI platform.

Perplexity

A pure web research provider (Sonar models) – it runs nothing locally and edits no files. Use it to research with sources, then implement the results with Claude or Codex.

Mistral (Vibe)

• Prerequisite: install the Vibe CLI once (a separate Python command line tool) – Liza then detects it automatically.
• Store a Mistral API key (create one at console.mistral.ai). Uses the Devstral 2 model as a local coding agent.

OpenCode additionally exists as a BYOK agent for Mistral models and self-hosted endpoints; it is currently hidden in the composer while the provider is being reworked.

Agents & skills

Settings → Personal → AI agents and skills – two tabs: Agents and Skills. Each has three sources:

Source	What it is
Write your own	You phrase the instruction yourself.
Provided	Maintained centrally and released for your organization. The AI finds them on its own; in the chat you can pick one deliberately.
Get from GitHub	Import from a public repository – see below.

Agents

Tailored agents – e.g. a "PR reviewer". They appear in the Claude Code tab as a selectable pill. Per agent you define:

• Name – e.g. "PR reviewer".
• Instructions / role – what the agent should do (and in which order).
• Steps (optional) – enforce a fixed order via "+ Step".
• AI – which provider runs the agent.
• Access – how much the agent may do:
  • Research / read-only – read and research only
  • Read-only + Liza – read + report back to Liza (check off / summary)
  • Coding (full access) – edit files and run commands
• Model – Default, Sonnet, Opus or Haiku.

With "Create with AI" you describe in one sentence what the agent should do and Liza drafts instructions and steps. Under "Examples" there are templates (PR reviewer, bug fixer, test author, refactoring, researcher).

read-only is not a hard lock with every provider

With Mistral, "read-only" is only a hint to the AI – the agent could still write. Reliably read-only works with Claude.

Skills

Skills are small instructions Claude pulls in automatically for matching work – e.g. "item report format" or "PR review checklist". Per skill you define:

• Name
• When to use – this is what Claude decides on
• Instructions – format, steps, rules

Your own skills apply on every Claude run without you picking them. In the web chat, the skills you chose for that conversation via "Apply skill" come along as well.

Not with every provider

Skills are pulled in by Claude. Mistral and Codex do not know this mechanism.

Sharing

Every skill has a share setting:

Setting	Effect
Only me	Applies in your AI sessions alone. The default.
Team	Applies in the AI sessions of every member of a team you are in.
Whole organization	Applies to everyone. Admins only.

A shared skill sits in every AI session of its recipients – which is why it is restricted: it is an instruction to their AI, not an offer. Anyone who does not want a shared skill can switch it off for themselves via "Switch off for me", without taking it away from others.

Get from GitHub

Plenty of ready-made skills and agents live in public repositories. Via "Get from GitHub" (in the menu next to New skill or New agent) you paste the address of a public repo, Liza reads it and shows what it found. With Adopt it becomes yours: private, editable as usual, and with no link back to the origin.

What is detected:

	What is looked for
Skill	a SKILL.md – along with the extra files that belong to it
Agent	a .md with agent front matter, usually agents/… or .claude/agents/…; a direct link to a single file works too

Three things matter here:

• Read the instruction before you adopt it. A skill is an instruction to your AI – afterwards it does what the text says. Liza shows you the full text and, where present, the repository's license.
• An agent arrives as "read-only". Tool permissions the repository asks for are not carried over; you set the access level yourself afterwards.
• Large repositories are not taken wholesale. Some collections hold hundreds of files; what is adopted is the instruction with its extra files, not the whole repository.

Connection to Liza & working directory

Two settings apply to all agents:

• Liza MCP key (optional): lets the agent check off tasks in Liza and post summaries. Leave empty = no write access to Liza. See Liza access for the key.
• Default working directory: the directory the agent works in. Applies to new lists; per task it can be overridden in the Claude Code tab.

"Set up Claude Code once — the connection, the Liza key and the default directory then apply in every list."

Local dev servers ("Servers")

The desktop app can start and supervise a project's server processes – database, backend, web client and so on. They are described in a .liza-run.json in the project ("environment"); individual dev servers can also be started right from the server box on a task.

Per process you see status, address and log (searchable, optionally errors and warnings only) and can stop, restart, open in the browser or add a second instance. Liza also warns about broken environments (duplicate ports, unknown dependencies, cycles).

Desktop only

Local server processes exist only in the desktop app.

Usage

Open the Claude Code tab (globally or inside a list), set the provider and working directory if needed, and write your job. On a single task you can also use the AI session with its job queue. Jobs can also be run on a schedule.