Universal autonomous AI worker
Fathom is a universal AI worker you self-host: it runs outreach, does deep research, writes code, operates a computer, and works on a schedule — with durable memory, governance, and human review built into one runtime.
Rust · one binary · memory · governance · schedules
00Autonomous Workspace
Complete multi-agent execution workspace: sub-agent swarm dispatch, live ARIA DOM browser control, and real-time CRM pipeline sync.
Fig 1.0 — Fathom Command Center: Swarm coordinator breaking objectives into parallel worker tasks, inspecting live Playwright browser viewport, and executing verified CRM writes.
Workers connect to your stack
01What your worker does
One goal in — finished work out. The same worker runs outreach, research, code and operations, switching roles as the task demands.
Find prospects across the web, verify every email through a five-gate pipeline, draft personalized sequences, and sync approved contacts to your CRM.
Seven search backends fused into one query, then fetch, parse, cross-reference and synthesize findings into a sourced brief.
Write and run code, inspect repositories, drive a browser and operate an optional connected computer — inside explicit tool boundaries.
Durable background jobs survive restarts, repeat on a schedule, and resume interrupted work exactly where it stopped.
Hybrid vector + BM25 search and an entity graph keep every finding across sessions — scanned for secrets before it lands.
Every bot in the sidebar is a real agent — Claude, Codex, or Grok. Chat like contacts, grant cloud computers, approve actions inline.
Every finished run lands where your team already looks — webhook, email or Telegram. One failing channel never blocks the rest.
02Agent types
Each sub-agent has a focused role and its own behavior. Coordinators decompose, fan out, reflect, and synthesize.
03How it works
Describe an outcome in plain language. The coordinator turns it into durable sub-tasks.
Workers delegate, use configured tools, and coordinate research, code, computer, or operations work.
Findings persist in memory; policies and approval gates make side effects reviewable.
Return a report, code change, approved update, or notification through the interface you choose.
04Performance
Real benchmarks from the release build — parallel tool execution spreads CPU across cores and ships verified results with minimal dispatch overhead.
See all benchmarks →05The full toolbox
Web, files, shell, code, browser, vision, Git, and verification; computer tools appear when the optional service is configured.
Vector + BM25 hybrid search, entity graph, and worker findings that persist across sessions.
Durable background jobs can continue across restarts and scheduled runs.
Approval gates, policy hooks, protected surfaces, and secret scanning keep side effects reviewable.
Connect external tools over stdio or HTTP, and expose Fathom tools through mcp-serve.
Native Tauri & GPUI desktop with ⌘K palette, interactive canvas click forwarding, full-screen voice calls, TTS, and Next.js dashboard.
Use built-in personas or define your own TOML worker profile.
Run repeatable work on a schedule and notify operators when configured.
Inspect sessions, replay events, and export Prometheus-compatible metrics.
Webhook, email, and Telegram fan-out alerts when configured and triggered.
06Safety & Reliability
Every agent action passes through multiple safety layers — from doom-loop detection to secret scanning to human approval gates.
Stuck agents repeating the same tool call are detected and stopped automatically — no runaway costs.
Side-effect tools like save_contacts and git_push wait for explicit human approval before executing.
API keys, passwords and tokens are blocked before they can enter the long-term memory store.
Prompt injection attempts in web-fetched content are detected and neutralized before they reach the model.
Crashed or interrupted sessions are detected and reconstructed — unfinished sub-tasks continue from where they stopped.
Safety configs, memory databases and own source code are guarded from self-modifying agent actions.
07One live run
The same honest path every time: goal → plan → tools → verification → approval → memory → delivery. Every stage is visible, every number is real.
08Deep work
A fathom is a nautical measure of depth. Ours is measured in gates and backends: seven search engines deep, five verification gates deep, every session replayable to the bottom.
1 fathom = 1.829 m · depth of work
Strategic Systems Whitepaper
Explore the definitive architectural and economic breakdown of autonomous AI employees: ~0.75 ms tool dispatch, 94 µs memory assimilation, 15.4 MB RAM footprint, and unlimited neural compute economics.
Self-hosted and reviewable. No vendor lock-in. Give your workers access to the tools, memory, and schedules your team controls.
fathom run "your task here" · docs · request access
09FAQ
Fathom is a universal autonomous worker runtime in Rust: it plans goals, coordinates sub-agents, uses code and computer tools, remembers context, and keeps side effects behind governance gates — not a single-session chat context window.
Seven backends — Linkup, Exa, Tavily, Serper, Brave, Parallel.ai and DuckDuckGo — fused with hybrid and smart modes and RRF ranking, so one query spans the whole web, not just the first page.
Fathom runs on servers we provision and manage for you. Sessions, contacts and memory live in your isolated environment. Contact us at [email protected] for details.
Use the CLI and TUI, or expose a configured runtime through HTTP/SSE, AG-UI compatibility, and MCP. The available surface depends on how you deploy and configure Fathom.