Agent Tools

The agent has access to built-in tools from the Claude Agent SDK plus 21 custom tools provided by Starpod.

Built-in Tools

These come from the agent-sdk and provide core capabilities:

Tool	Description
Read	Read a file from disk
Write	Write content to a file
Edit	Make targeted edits to a file
Bash	Execute shell commands
Glob	Find files by pattern
Grep	Search file contents with regex
WebSearch	Search the web
WebFetch	Fetch a URL

The Bash tool supports run_in_background: true for long-running processes (servers, watchers, etc.) that shouldn't block the agent.

Custom Tools

Memory

Tool	Input	Description
MemorySearch	query, limit	Full-text search across all memory files
MemoryWrite	file, content	Write or update a memory/knowledge file
MemoryAppendDaily	text	Append to today's daily log

Environment

Tool	Input	Description
EnvGet	key	Look up an environment variable by key

Environment variables from the vault are injected into the process at serve time, making them available both through EnvGet and as real env vars in Bash/SSH commands. The system prompt lists which vars are available. System keys (LLM provider keys, service tokens) are blocked by EnvGet and stripped from Bash child processes.

Files

Tool	Input	Description
FileRead	path	Read a file from the agent's filesystem sandbox
FileWrite	path, content	Write a file to the agent's filesystem sandbox
FileList	path (optional)	List files and directories in the sandbox
FileDelete	path	Delete a file from the sandbox
Attach	path	Attach a sandbox file for delivery to the user (see below)

Attach

The Attach tool sends a file to the user through their current channel. After the agent generates or locates a file in its sandbox, calling Attach queues it for delivery:

• Web UI — delivered as a WebSocket attachment message; images render inline, other files show as download links.
• Telegram — images sent via send_photo, everything else via send_document.
• CLI — files remain in the sandbox (already accessible on disk).

The tool validates sandbox paths the same way FileRead does (no .. traversal, no absolute paths, no .starpod/ access). Files must be under 20 MB. MIME type is inferred from the file extension.

Multiple files can be attached in a single turn — they accumulate and are delivered after the agent finishes responding.

Skills

Tool	Input	Description
SkillActivate	name	Load a skill's full instructions into context
SkillCreate	name, description, body	Create a new AgentSkills-compatible skill
SkillUpdate	name, description, body	Update an existing skill's description and instructions
SkillDelete	name	Delete a skill
SkillList	—	List all active skills

Scheduling

Tool	Input	Description
CronAdd	name, prompt, schedule, delete_after_run, max_retries, timeout_secs, session_mode	Create a scheduled job
CronList	—	List all jobs with next run times
CronRemove	name	Remove a job
CronRuns	name, limit	View execution history
CronRun	name	Immediately execute a cron job (manual trigger)
CronUpdate	name, prompt, enabled, max_retries, timeout_secs, session_mode	Update properties of an existing job

Heartbeat

Tool	Input	Description
HeartbeatWake	mode ("now" or "next"), message	Wake the heartbeat system outside its normal cycle

CronAdd Schedule Format

The schedule parameter accepts three formats:

Interval

{
  "kind": "interval",
  "every_ms": 3600000
}

Cron

{
  "kind": "cron",
  "expr": "0 0 9 * * MON-FRI"
}

One-shot

{
  "kind": "one_shot",
  "at": "2026-03-14T15:00:00Z"
}

Tool Execution

The agent-sdk drives the agentic loop:

1. Claude receives the system prompt + conversation history
2. Claude decides to call one or more tools
3. The SDK executes the tools and feeds results back to Claude
4. Claude decides whether to call more tools or respond
5. Repeats up to max_turns (default: 30)

Custom tools are handled by an external tool handler registered on the Options builder. When Claude calls a tool, the handler checks if it matches a custom tool name. If it does, the handler executes it and returns the result. Otherwise, the SDK's built-in handler runs it.