Tools Reference

Search the internet via the configured backend (Linkup/Exa/Tavily/Serper/Brave/Parallel/DuckDuckGo).

Returns: List of results (title, URL, snippet).

Download a page and convert HTML to readable text.

Returns: Page text (truncated to 50K chars).

Crawl a website breadth-first from a seed URL.

Returns: Per-page title + extracted text.

Parse an RSS 2.0, Atom, or sitemap.xml feed into a clean item list.

Returns: Feed items (title, link, description, pub_date).

Read a file with optional partial reading support.

Note: Tracks read files for read-before-write validation.

Write a file (creates directories automatically).

Note: Per-path locking, updates file history.

Find/replace edit (surgical replacement).

Validation: read-before-write, staleness check, size guard.

Search files by glob pattern.

Search file contents with regex. Uses ripgrep when available.

Execute a bash command.

Protection: Blocks destructive commands (rm -rf /, mkfs, fork bombs).

Find code definitions (functions, classes, structs, traits, methods) in a file or directory. Supports Rust, Python, JS/TS, Go, Ruby, Java/Kotlin, C/C++/C#, PHP.

Returns: List of definitions with name, kind, file, line.

Build a compact map of a codebase: files by language plus their top symbols.

Returns: File tree with language tags and top-level symbols per file.

Execute Python code and return stdout/stderr. Uses system python3.

Returns: stdout, stderr, exit_code.

Execute Node.js code and return stdout/stderr. Uses system node.

Returns: stdout, stderr, exit_code.

Query the Language Server Protocol for code intelligence: document symbols, go-to-definition, find references, hover info, and workspace symbol search. Auto-detects LSP server from project files.

Returns: LSP response (symbols, locations, hover info depending on action).

Open a URL in the browser.

Take a screenshot of the current page.

Returns: base64 image in metadata.base64.

Click an element by CSS selector.

Type text into an element.

Extract text content from page (with JS rendering).

Returns: Page text (truncated to 50K).

Analyze an image via vision model (qwen-vl-max).

Config: PARALLEL_VISION_API_BASE, PARALLEL_VISION_API_KEY, PARALLEL_VISION_MODEL.

Show working tree status.

Show changes.

Commit history.

Add files to staging.

Create a commit.

Push to remote.

Extract text from a PDF file.

Returns: Text + page count. Supports ToUnicode CMaps, encryption (empty password).

Extract structured data from an HTML page or file using CSS selectors. Modes: texts, html, attr, links, tables.

Returns: Extracted elements matching the selector.

Query a JSON document (API response, file, or inline value) with a dot path and return selected values.

Returns: Matching values at the specified path.

High-level lead search. Orchestrates business directories → corporate sites → social networks → contact extraction → deduplication.

Search business directories (2GIS, Google Maps, Yandex Maps, Yellow Pages).

Returns: Name, category, address, phone, website, email, rating.

Search people/companies on social networks (Twitter, Telegram, LinkedIn).

Parse corporate website: company info, team, contacts.

Returns: Company name, description, industry, size, contacts, team page URL, social profiles.

Search news/mentions (Serper News or Google News RSS).

Returns: Title, URL, source, date, snippet + extracted persons/companies.

Verify email: syntax, MX records, disposable, role-based checks.

Returns: is_valid_syntax, domain_exists, mx_records, is_disposable, is_role_based, confidence.

Generate likely corporate email addresses for a person from standard name permutations at a domain. Infers the company’s email pattern from known addresses. Every candidate is verified.

Returns: List of candidate emails with verification status and confidence scores.

Validate and normalize phone number (E.164).

Returns: Normalized (E.164), country_code, country_name, is_valid, is_mobile.

Check if a social media profile exists.

Returns: exists, platform, username, name, followers.

Enrich company data.

Returns: Industry, size, revenue, founded, headquarters, description, technologies.

Enrich person data.

Returns: Title, company, linkedin, twitter, email, phone, location, bio.

Extract structured contact information (emails, phone numbers, social profiles, people, companies) from text, HTML, or a URL. Built for OSINT and lead-generation research.

Returns: Emails, phones (E.164), social profiles, persons, companies — each with confidence and source.

Save harvested contacts into the persistent contact database (contacts.db), and optionally push to CRM (amoCRM/Bitrix24/HubSpot) when configured.

Returns: saved count, merged count, skipped count.

Spawn a sub-agent to handle a specific research task. Available roles: researcher, analyst, verifier, writer. Sub-agent runs autonomously with its own context.

Limit: Nesting depth (max_depth).

Ask the human operator a question and wait for the answer. Use ONLY when genuinely blocked. Max 500 chars.

Returns: The operator’s answer text.

Load the full instructions of a skill by name. Skills are reusable playbooks from ~/.fathom/skills.

Returns: Skill instructions as text.

Shared session ledger for coordination between agents. Actions: read (return current content), append (add one line). Prevents duplicate work across parallel agents.

Returns: Current scratchpad content (for read) or confirmation (for append).

Undo recent file edits made during this session by rewinding to an earlier file-history checkpoint.

↩ Confirmation of the undo operation with list of reverted files

Manage persistent memory entries stored in MEMORY.md and USER.md under ~/.fathom/memory/. Actions: add, replace, remove, or batch.

Returns: Memory entries or operation result.

Store one or more facts into long-term semantic memory. Facts are deduplicated, reconciled, and near-duplicates consolidated. Secrets are rejected automatically.

Returns: Absorb report (created, duplicates, superseded counts).

Search long-term semantic memory. Combines vector similarity with keyword (BM25) matching and recency decay.

Returns: Ranked memories with scores.

Load a deterministic context digest from long-term memory for a topic: most relevant memories, open TODOs, recently added facts.

Returns: Formatted digest text for system prompt injection.

Raise (or lower) the importance of a stored memory. Importance influences ranking under tight score margins and future digests.

Returns: Updated memory with new importance.

Add a typed edge between two stored memories. Edge types: related_to, supersedes, contradicts, implements, extends, references.

Returns: Created edge confirmation.

Entity knowledge graph for persons, companies, and locations. Stores entities and typed relations; answers multi-hop queries.

Returns: Entities and relations matching the query.