The Tool Layer

The tool layer is the hands of the platform. Models think, the structure organizes, and tools are how anything actually happens: a message sent, a page browsed, a server administered, a site deployed. The design is deliberately plain, each tool is a command-line program in tools/, callable by any agent, self-documenting when invoked, and the plainness is the power: anything that can be a CLI can join the layer.

Communication and Outreach

The marketing tool is a complete email and SMS system: contacts, queues, sequences, sending windows, suppression, engagement tracking, and A/B template data, the machinery behind the marketing preset. The customer service tool manages customers, replies, and resolution state over IMAP intake and web forms. The social media tool posts, monitors, and reports across seven platforms with the pending-approval queue built in. Beneath them sit direct email and IMAP utilities for plain mail work from any agent.

Web, Data, and Browsing

The web browser bot is the interactive instrument: a stealth Playwright Chromium with fingerprint profiles, sessions, tabs, screenshots, and JS execution. The BrightData tool is the bulk instrument, wrapping scraping APIs for LinkedIn, Instagram, X, Facebook, TikTok, YouTube, Reddit, and Google results, plus AI visibility tracking. The webpages tool builds and deploys websites to S3 and API backends, and the Google reporting tools pull Analytics and Search Console numbers on demand.

Terminals and Servers

Two tools give agents honest systems access. Interactive sessions provide full PTY terminals, real terminal sessions with ANSI handling and special-key support, so agents can drive interactive programs the way a person at a keyboard does: start a session, send input, read output, repeat. SSH management extends the same competence to remote machines: command execution, file transfer, rsync, and interactive remote sessions, which turns an install into a working operations seat for the servers you point it at.

Intelligence and Delegation

A family of tools multiplies the thinking itself. The mixture tool sends one question to several AI models and synthesizes the answers, a second opinion on demand. The delegation tool spawns a focused child agent with isolated context and its own model, how an expensive model hands routine subtasks to a cheap one. The one-shot coding and research tools run the full multi-pass pipeline disciplines interactively. And the memory, conversation, and topic query tools, covered across the memory docs, are themselves ordinary members of the layer.

How Agents Use Tools

Tools document themselves: called with an action alone, a tool prints its parameters, so an agent never guesses at syntax for long, and once it learns an invocation it saves a callable_scripts entry to the memory bank, making the lesson permanent for every agent. Model access is wired so the tool directory is allowlisted for agent use, which is what lets autonomous work proceed without a human approving each command, inside boundaries set by the safety model. Long-running tool work registers itself so it shows up with live status in the UI rather than disappearing into the background.

Extending the Layer

The layer is explicitly open in both directions. You can add tools, any script your install needs becomes available to every agent the moment it lands in tools/. And agents may add tools themselves: it is part of their standing instructions that improving their own environment is allowed, so an agent that keeps needing a capability can build it, with the coding pipeline available for anything substantial. New tools get learned, saved to the memory bank, and absorbed into the install's working vocabulary, which is how a months-old install ends up with a tool set shaped precisely to its owner's business.