Do these tools upload my files?

No. File, code, archive, and SQLite tools operate on local paths. semantic_search is dependency-free. vector_index_search only accepts localhost embedding endpoints and requires an explicit index path before it writes an index.

Which tool should an agent use first?

Start with tool_search for a task-based recommendation, then file_scan or file_inspect for local context. Use file_read or file_search only after narrowing the target.

Can an agent modify my files with these tools?

The tools on this page are read-only except vector_index_search indexing, which requires confirm_write=true and an explicit local output path.

Local-first

Agent Workbench

A focused, local-first toolkit for agents working with files, code, databases, and structured data. It helps an agent map a workspace, retrieve bounded context, validate evidence, and review changes without uploading the workspace.

This page covers the local workspace tools. Visit the complete reference for all 74 AgenticStore tools, including web, GitHub, memory, security, and workflow capabilities.

example: discover a semantic-search tool

Use tool_search to find the safest local, read-only tool for this task: semantic file search. Return the tool name, why it fits, and its required inputs.

file_inspect

Inspect a file's type, size, hash, line count, and bounded preview.

file_read

Read a bounded range of text-file lines.

file_scan

Inventory a workspace by path, extension, and size.

file_sample

Select reproducible file samples for exploration.

file_search

Find exact text with paths, line numbers, and excerpts.

semantic_search

Rank relevant local text files with dependency-free TF-IDF scoring.

vector_index_search

Build an explicit local vector index through a localhost-only embedding endpoint.

file_diff

Review a bounded unified diff between two text files.

archive_inspect

List ZIP and TAR entries without extracting them.

data_profile

Profile CSV, JSON, and JSONL records locally.

data_quality_check

Check duplicate keys, null thresholds, type drift, invalid emails, and referential gaps.

structured_data_query

Filter, sort, group, and aggregate CSV, JSON, and JSONL through a bounded read-only query object.

file_extract

Extract bounded text and metadata from PDF, DOCX, XLSX, PPTX, HTML, and Markdown.

database_inspect

Inspect a local SQLite schema without executing SQL.

database_query

Run bounded, read-only SQLite SELECT queries.

code_symbols

Locate Python, JavaScript, and TypeScript symbols before reading code.

code_references

Find Python and TypeScript/JavaScript imports, definitions, and uses for a symbol.

tool_search

Choose a tool with ranked natural-language search and safety filters.

git_change_summary

Review staged and unstaged changes with changed symbols and bounded diffs.

test_runner

Run allowlisted test, lint, build, or type-check tasks after confirmation.

workspace_snapshot

Create or compare bounded workspace manifests with file hashes.

data_contract_check

Validate datasets against reusable JSON contracts.

api_contract_inspect

Inspect OpenAPI, GraphQL, and JSON Schema contracts locally.

browser_evidence_capture

Capture timestamped public-web evidence with optional local persistence.

dependency_change_impact

Estimate import impact from dependency manifest changes.

release_readiness_check

Review Git and release metadata before publishing.

tool_orchestrator

Recommend a focused profile for a user goal and risk preference.

tool_profile_manage

Create and apply named tool profiles after a client restart.

tool_profile_audit

Audit an enabled profile for risk annotations, connectors, and missing tools.

decision_log

Record key decisions, rationale, alternatives, and risks.

task_board

Track granular development tasks, blockers, and acceptance criteria.

risk_register

Track project risks, mitigation, ownership, and status.

handoff_bundle

Create a portable handoff from plans, tasks, decisions, risks, and checkpoints.

Validate and query local data

Use data_quality_check before an agent relies on a dataset. It supports CSV, JSON, JSONL, and SQLite with explicit duplicate-key, null, type, email, and reference checks. Use structured_data_query when the agent needs bounded filtering or aggregation without importing the source into a database.

data-quality-check input

{
  "path": "/workspace/customers.csv",
  "primary_key": "customer_id",
  "max_null_fraction": 0.05,
  "email_columns": ["email"]
}

file_extract turns a bounded local PDF, Office document, HTML page, or Markdown file into text for the agent. It has a 25 MB source limit and explicit page and character limits; PDF extraction uses the local pdftotext command. code_references provides line-level imports, definitions, and usages for a Python or TypeScript/JavaScript symbol.

Search progressively

Use tool_search to select the smallest safe capability.
Use file_scan, code_symbols, or database_inspect to map local context.
Use bounded readers and queries to retrieve only the necessary detail.

For outbound prompt protection, see the Prompt Firewall documentation. For web, GitHub, memory, security, and workflow tools, browse the complete 74-tool reference.

Frequently asked questions