Sandbox Environment

What’s inside the agent sandbox — tools, CLI, and capabilities

Every agent run executes in an isolated container (sandbox) with a pre-configured environment. The sandbox is created fresh for each run and destroyed after completion.

What’s pre-installed

Component	Version	Description
AI Agent	Latest	AI coding agent with tool use, file editing, and web fetching
casedev CLI	v0.1.1+	Pre-authenticated CLI for all Case.dev services
Node.js	22	Runtime for scripts and tools
curl	System	HTTP requests

The casedev CLI

The CLI is pre-authenticated with a scoped API key — no login required. The agent can call any Case.dev service:

Vault operations

Vault operations

# List vaults
casedev vault list

# Search a vault
casedev search vault "force majeure clauses" --vault vault_abc

# Upload a file
casedev vault upload ./report.md --vault vault_abc

# Download files
casedev vault download --vault vault_abc --object obj_xyz

Legal research

Legal research

# Search case law
casedev search legal "employment discrimination ADA"

# Deep research with multiple queries
casedev search legal "trade secret misappropriation" --deep

# Search patents
casedev search patent "natural language processing" --assignee "Google"

# Search similar cases
casedev search cases --url "https://case-url" --after 2020-01-01

Web search

Web search

# Web search
casedev search web "recent SCOTUS rulings on digital privacy"

# Fetch and summarize a webpage
casedev search webfetch https://example.com/ruling.pdf

Other services

Other services

# OCR a document
casedev ocr process --document-url "https://..."

# Transcribe audio
casedev transcribe run --object obj_xyz --vault vault_abc

# Call any API endpoint directly
casedev api GET /vault/vault_abc
casedev api POST /llm/v1/chat/completions --body '{"model": "...", "messages": [...]}'

Run casedev --help inside the sandbox to see all available commands. The CLI covers every Case.dev service.

Agent tools

The AI agent has access to these tools for working within the sandbox:

Tool	Description
bash	Execute shell commands (including the casedev CLI)
read	Read files from the sandbox filesystem
write	Create or overwrite files
edit	Make targeted edits to existing files
glob	Find files by pattern
grep	Search file contents
webfetch	Fetch and parse web pages
todowrite	Track multi-step task progress

The agent uses these tools autonomously to complete your prompt. For example, when asked to “search a vault and compile a report,” it will:

1. Run casedev vault list via bash to find vaults
2. Run casedev search vault to find relevant content
3. Use write to create a markdown report
4. Run casedev vault upload to upload the report
5. Run casedev api to get a presigned download URL

Sandbox resources

Resource	Default	Configurable
CPU	2 cores	Yes, via sandbox.cpu
Memory	2048 MiB	Yes, via sandbox.memoryMiB
Timeout	30 minutes	No
Disk	Ephemeral	No
Network	Full outbound	No

Configure resources when creating the agent:

TypeScript

const agent = await client.agents.create({
  name: 'Heavy Processor',
  instructions: 'Process large document sets.',
  sandbox: {
    cpu: 4,
    memoryMiB: 4096,
  },
})

Filesystem

The sandbox starts in /workspace with a clean filesystem. The agent can create, read, and modify files freely. Key paths:

Path	Purpose
/workspace	Working directory
/workspace/agent.json	Agent configuration (auto-generated)
/usr/local/bin/casedev	CLI binary

The sandbox filesystem is ephemeral. Files are destroyed when the run completes. To persist results, upload them to a vault using casedev vault upload.

Security model

Each sandbox is fully isolated:

• Dedicated container — no shared state between runs
• Scoped API key — the agent’s key only has access to authorized vaults and services
• Network isolation — outbound access only, no inbound connections
• Auto-cleanup — sandbox is destroyed after run completes or times out

The agent’s API key is automatically revoked if the agent is deleted.

Limitations

Limitation	Details
No inbound networking	The sandbox cannot receive incoming connections
30-minute timeout	Long-running tasks must complete within this window
No GPU	CPU-only environment
No persistent storage	Use vaults for any data that needs to survive the run
Single session	One agent session per run