Running Code with Execute Mode

Execute mode is Zaru's single-shot code execution mode. You describe what you want to compute in plain English, optionally provide typed inputs, and Zaru generates the code, runs it in a secure isolated container, and returns the result directly in chat.

Prerequisites

The Execute mode pipeline requires the builtin-intent-to-execution workflow and its supporting agents (intent-executor-discovery-agent and intent-result-formatter-agent). These are automatically deployed when deploy_builtins is enabled in your node configuration.

Switching to Execute Mode

Use the mode selector at the top of the conversation to choose Execute before sending your first message. The mode is locked once the first message is sent.

The mode selector is available on all tiers. Execute mode requires no workflow authoring, no agent deployment, and no volume configuration.

Context Panel in Execute Mode

When you expand the Context Panel (the sidebar icon in the top-right corner of the chat), you see two tabs: Workflows and Tools. These let you browse and interact with available resources without going through the chat input:

• Workflows tab — lists deployed workflows. Click Execute on any workflow to open an inline parameter form and run it directly; the result appears inside the panel. Click Send to Chat to have Zaru walk you through the workflow's inputs before running it.
• Tools tab — lists available tools in the same way, with direct execution or chat-assisted invocation.

Motivating Example

Type your intent into the chat:

add 2 numbers together

With inputs:

{ "a": 1, "b": 1 }

Zaru returns:

Result: 2

The Glass Lab panel shows a four-stage progress view as the pipeline runs: Discover / Generate Agent → Generate Code → Execute → Format Result.

How It Works

Execute mode runs a built-in four-state pipeline behind the scenes. The Zaru client uses aegis.workflow.wait (aliased as aegis.execute.wait) to block server-side until the pipeline reaches a terminal state, rather than polling aegis.execute.status in a loop.

Supported Languages

Python is used when no language is specified. To request a specific language, mention it in your intent: "write a JavaScript function that…"

Providing Inputs

Pass typed inputs as a JSON object alongside your intent. Scalar values are embedded directly into the generated code — you do not need to write any parameter-passing logic.

Scalar inputs:

{ "a": 42, "b": 7 }

String inputs:

{ "text": "hello world", "delimiter": " " }

Array inputs:

{ "numbers": [1, 2, 3, 4, 5] }

Object inputs:

{ "config": { "width": 800, "height": 600 } }

For complex types (arrays, nested objects), the inputs are also written to a params.json file in the workspace. The generated code can read from this file if direct embedding is not appropriate.

Workspace and Storage

By default, Execute mode uses an ephemeral workspace: a temporary volume created at pipeline start and destroyed immediately when execution finishes. No data persists between runs.

Pro and higher tiers can pass a volume_id to write results to a persistent volume:

{ "intent": "generate a CSV report of monthly totals", "volume_id": "vol-abc123" }

Tier Limits

Security

Code executes inside a container with the following restrictions applied unconditionally — these are not configuration options:

• No internet access — the container has network: none
• Read-only workspace — the generated file is mounted read-only; the container cannot modify it
• Read-only root filesystem — no writes to any path on the host
• Unprivileged user — runs as nobody (uid 65534); no privilege escalation
• No SEAL session — no access to AEGIS tools or the orchestrator API
• No LLM access — no calls to any language model from inside the container

Only stdout, stderr (each capped at 1 MB), and the exit code cross the container boundary. The AEGIS membrane ensures your code cannot access other users' data or the host system.

Common Use Cases

• Quick math or unit conversions
• Data transformation and reshaping
• String processing and parsing
• Algorithm prototyping
• Sorting, filtering, and aggregating arrays
• File generation (Pro and higher, using a persistent volume)

Limitations

• No internet access inside the execution container
• Standard library only — pip install, npm install, and similar package managers are not available
• Execution time is capped per tier (see tier limits above)
• Code must be self-contained — no interactive stdin
• Free tier cannot use persistent volumes

Related

• Zaru Integration — conversation modes, tier routing, and the Glass Laboratory UX
• Configuring Storage — managing persistent workspace volumes (Pro and higher)