Loading ToolBench
A collection of three Claude Agent SDK examples demonstrating different agent patterns: account research agent, document pipeline agent with custom tools and hooks, and an inbox triage service with HTTP hosting.
This is a multi-example agent repository, not a traditional MCP server. Tools are scattered across Python and TypeScript examples without centralized registration. Custom tools (lookup_vendor, check_sow, check_rate, record_extraction) have reasonable descriptions (80-150 chars) and typed input schemas, but built-in SDK tools are documented minimally. No output schemas are visible. The repository demonstrates agent patterns but lacks the structured tool definitions expected of a production MCP server. STDIO transport is the only option, capping protocol readiness.
Built-in Claude Agent SDK tool for running shell commands (explicitly disallowed in example 01)
Built-in Claude Agent SDK tool for editing files (explicitly disallowed in example 01)
Built-in Claude Agent SDK tool for listing files matching a pattern
Built-in Claude Agent SDK tool for editing notebooks (explicitly disallowed in example 01)
Built-in Claude Agent SDK tool for reading files
Built-in Claude Agent SDK tool for delegating to subagents
Built-in Claude Agent SDK tool for fetching web pages
Built-in SDK tools (WebSearch, WebFetch, Read, Write, Glob, Bash, Edit, NotebookEdit, Task) lack visible input schemas and have minimal descriptions. These are inferred from SDK documentation, not explicitly defined in this codebase.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract structured data without knowing what fields to expect.
Custom tools lack error handling guidance. No recovery hints, retryability classification, or actionable error messages documented.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 36 | 2026-07-28+ | v2 |
Built-in Claude Agent SDK tool for searching the web
Built-in Claude Agent SDK tool for writing files
Validates billed rates against the contracted rate card for a specific SOW and role. Compares billed rate to contracted rate and flags discrepancies.
Validates a Statement of Work reference against the SOW register. Checks SOW format, existence, vendor match, status, and authorization ceiling. Returns detailed findings including remaining authorization and over-authorization warnings.
Looks up a vendor in the vendor master by name, performing slug matching and prefix matching. Returns vendor record with legal name, vendor ID, category, MSA status, and payment terms. Blocks vendors with non-active MSA status.
Records an extraction result to the extraction ledger. Appends a structured record of extracted invoice data with validation findings and review flags.
Destructive tools (Bash, Write, Edit, NotebookEdit) lack confirmation or dry-run patterns. No permission gates or audit trail documented.
Repository is a multi-example agent codebase, not a centralized MCP server. Tools are scattered across examples without unified registration, making it unsuitable as a production MCP service.