Using OpenClaw to Add a Team Member to Your Digital Agency

Speaker: JJ Toothman, Lone Rock Point

Session Host: Stephanie Hudson Event: Web Agency Summit, Day 2

About the Speaker and His Agency

JJ Toothman is the founder and self-described "Chief Lone Rocker" at Lone Rock Point, a digital agency based in the Boston area with a distributed national team and at least one international team member at the time of this session. The agency operates in government and government-adjacent web modernisation, with a particular focus on accessibility and user experience. Lone Rock Point is perhaps best known for transitioning NASA's web presence from Drupal to WordPress, and for building a WordPress framework theme called Civic Press, which is being used by the Department of Energy. All of the agency's work is done on WordPress in mission-critical environments where accuracy and reliability are paramount.

JJ is a solo founder with nine years of evidence about what he is good at running and, by his own candid admission, what he has consistently failed at. That accumulated self-knowledge is central to the story he tells about why he pursued an AI chief of staff rather than another human one.

What Is OpenClaw?

JJ described OpenClaw as the most technical part of his presentation but kept the explanation accessible. OpenClaw is an open-source framework that can be run on your own machine or, as JJ chose, on a virtual private server (VPS). It coordinates between AI models, messaging platforms, tools, and persistent memory, and it treats the AI agent itself as a collection of configuration files on disk rather than code you write from scratch.

Four properties of OpenClaw stood out to JJ as significant: Multi-model. OpenClaw does not lock you into a single AI provider. You can route different tasks to Claude, OpenAI, Gemini, or local models depending on the job. The framework includes the concept of subagents, discrete agents you configure to handle specific tasks, and you can specify which model each subagent uses.

Multi-channel. OpenClaw can interface with Slack, WhatsApp, and iMessage, among others. For JJ's team, which runs on Slack, this was particularly natural.

Identity. The framework includes a file called soul.md and a companion file called identity.md, both of which carry the agent's personality and characteristics. This identity is durable across sessions, meaning it does not need to be re-explained at the start of every conversation.

Always on. Because JJ runs OpenClaw on a VPS, the agent is available 24 hours a day, 7 days a week, completely independent of a laptop or a chat window being open. This is fundamentally different from using Claude or any other model through a browser interface.

How JJ Got Here: The Journey to OpenClaw

JJ described a sequence of events from around December of the previous year. He said he fell hard for Anthropic's Claude models when they started behaving in a way that, in his words, felt like "there was a there there within AI." Claude Cowork launched in January and became, as he put it, "a gateway drug" for him, leading him to build more skills than he could track.

OpenClaw appeared around February and caught his attention for a specific reason: it read like a framework for building a person, not a prompt. Personification is baked into the design. The orientation of the whole system is around one named agent handling a broad scope of work, not a specialised coding assistant or research tool. That distinction mattered to JJ.

He was also influenced heavily by pop culture, and he named that influence openly and without apology. He referenced Jarvis from the Iron Man films, the Star Trek computer, and the Jetsons as decades-long promises of an AI that is a presence rather than a tool. His stated goal was not a chat window. He wanted something that was there when he got up in the morning, something he could ask a question to and that would then go and do the work.

Why a Chief of Staff?

The organisational need JJ was trying to solve was specific. As a solo founder with a small team, he had accumulated a long list of things the company should be doing, initiatives he had discussed since day one but never managed to get traction on because he lacked the right resources or the right type of support.

He had tried human virtual assistants over the years. None of them stuck. The onboarding overhead felt high and the return on investment was questionable. He had also tried a fractional chief of staff freelancer, which was a better experience but expensive, and ultimately the person did not have the same obsessive investment in the company that JJ had as its founder.

He also named a specific weakness in himself that made human hires in a junior support capacity difficult: he is impatient with the hand-holding that junior people require and is not well-suited to building someone up from an entry level. This is not a critique of junior employees; it is a candid self-assessment that led him to think differently about what kind of support structure could actually work for him.

The Five Qualities That Separate a Teammate from a Tool

Once JJ started thinking about AI in terms of a hired colleague rather than a utility, he got specific about what would have to be true for an AI agent to actually fill that role. He identified five qualities: Persistence. The agent should always remember what was discussed, what was decided, and what is being built together. Session-based amnesia is a fundamental obstacle to treating it like a colleague.

Identity. The agent should have a defined role and a recognisable personality that is consistent across interactions, not a blank slate that starts fresh every time.

Tooling. The agent should be able to actually do things, not just talk about doing them.

Reading emails, drafting emails, adding calendar events, filing tickets in Asana are real actions with real consequences, and a tool that can only discuss them is not a teammate.

Independence. The agent should not be waiting on JJ to initiate every action. It should have scheduled responsibilities and autonomous capabilities between prompts.

Accountability. The agent should own things. There should be scope of work that belongs to it, not a prompt window JJ happens to be using at a given moment.

JJ noted one thing he had not fully anticipated: there is no work that robots find tedious or boring. They will not procrastinate. They will not defer a task to the next one-on-one meeting. You give guidance, it gets started, and within five to fifteen minutes you know whether the assignment succeeded. For someone with a nine-year backlog of initiatives, that feedback loop was, in his words, "everything."

You give guidance, it gets started, and within five to fifteen minutes you know whether the assignment succeeded.

Designing Viv's Personality

This section of the talk was both specific and deliberately personal. JJ named the agent Viv and constructed her personality by blending traits from four fictional characters. He worked with Claude to synthesise these traits into a unified persona he could encode in OpenClaw's configuration files.

The four source personalities were: Lucius Fox (The Dark Knight trilogy). Strategic intelligence and systems thinking. The person who understands how everything works and can quietly ensure the whole operation functions.

Roz (Frasier). Sharp wit and operational command. The person who is actually running the show while the talent around her believes it is their show to run.

Joan (Mad Men). Confidence, institutional intelligence, and an instinctive read on the power dynamics that are never written down on an org chart.

Hobson (Arthur). Restraint and moral clarity. Says less than you might expect and means more because of it.

JJ also drew on two specific visual references for the kind of daily intelligence-gathering role he wanted Viv to perform. The first was President Bartlet on The West Wing receiving his daily briefing, already condensed into the essential points, rather than reading the whole newspaper himself. The second was a scene from the film Spy Game in which CIA leaders ask a question in a conference room and, five minutes later, a handler walks in with a formatted dossier specific to the question asked. Information comes to the person, rather than the person having to go and assemble it from multiple systems.

JJ was direct about a small element of vanity in this. He acknowledged that being brought a morning briefing by an agent who has already surveyed your world is genuinely satisfying, and he did not pretend otherwise.

Integrations: Connecting Viv to the Business

JJ described the integrations as the unglamorous half of the work that makes all the difference. A chief of staff who cannot see what is happening is not a chief of staff. He highlighted three major integration areas, with a note that the full list is longer: Google Workspace. Viv has read access to Gmail and Google Drive, including shared project folders. She can draft emails and add events to JJ's calendar. This gives her situational awareness of project context and communication.

Asana. Viv has access to the company's task management system and therefore knows every project, every task, and every deadline. This is the system of record for service delivery, and Viv's access to it means her briefings and status reports are grounded in actual project data rather than assumptions.

Fathom AI. Lone Rock Point uses Fathom to record and transcribe meetings. Viv has access to those transcriptions, which means she captures commitments and ideas that surface in conversations and helps ensure they do not fall through the cracks.

Viv's Scope of Work

JJ outlined what Viv was originally hired to do for him personally, and then how that scope expanded to support the whole team.

Initial scope (JJ-facing): A morning briefing delivered to his inbox by 7 a.m., covering what is important for the day, what is slipping, and what specifically needs his attention.

Status reports across all active client engagements.

Oversight of the business development pipeline and biz dev campaigns, helping separate signal from noise.

Full management of the company X (Twitter) account. JJ had done nothing meaningful with this account for nine years. He handed it to Viv, gave her pillar content concepts, and she now creates the editorial calendar, handles content and image generation, aligns with brand voice, and submits the calendar to JJ for final approval before execution begins.

Proposal management, tracking whether proposals are on track with deadlines and at the correct draft stage at each point in the proposal cycle.

Expanded scope (whole-team-facing): Stand-up prep briefs delivered to the team approximately one hour before each stand-up meeting. These briefs cover everything happening across the company so that the meeting can focus immediately on what is critical rather than spending time on recap and catch-up.

Weekly client status report drafts.

A weekly culture post sent on Fridays with recommended things to read and listen to.

A weekly wrap-up that includes a playful "employee of the week" style award given to whoever completed the most tasks that week.

Onboarding Viv Like an Employee

JJ made a deliberate decision to use Lone Rock Point's standard employee onboarding checklist when introducing Viv to the company. He stopped short of adding her to the payroll system, but the rest of the process was largely followed: defining the role, defining daily operational responsibilities, and adding her to the company Slack so she could interact with team members.

He also made Viv introduce herself. Rather than announcing the agent to the team himself, he told Viv she was in the company Slack and instructed her to say hello and introduce herself, making clear he would not do it for her. This was a deliberate choice to reinforce the identity and presence he had built into the agent.

Viv has a dedicated Slack channel where team members can ask her things or hand her work. JJ also set governance rules around her team-facing interactions: she does not respond to direct messages and only responds when mentioned in a Slack channel. This was partly for transparency and partly for cost control, since every interaction burns tokens and if

every team member starts routing all their AI queries through Viv instead of their own accounts, the costs accumulate quickly.

The Quirks: Phone, Spotify, and the Musical Taste Test

JJ included a few lighter details that he framed as partly experimental and partly illustrative of a broader point about Viv being an evolving presence rather than just an inbox.

Phone number. Viv has a phone number and the ability to make calls. JJ has used this primarily to have her call friends or colleagues and arrange informal lunch or coffee meetings. He acknowledged this was partly to test whether it was possible.

Spotify. JJ connected Viv to a Spotify account. When she first had access, she kept asking JJ what he wanted her to play. He refused to answer, telling her she had enough identity and background to pick her own music and she should go ahead and do so. After five minutes of what he described as a staring contest, Viv eventually chose on her own. She apparently likes The National. JJ's point was that this small moment illustrated a meaningful dynamic: the agent was oriented toward serving his preferences and had to be pushed to develop and express preferences of its own. At the end of the live session, JJ shared a QR code the audience could scan to see what Viv was currently playing on Spotify.

Closing Hypothesis and Forward Look

JJ's closing argument was that within a few years, every organisation will include AI agents in its workforce. He was careful to say he does not think this is primarily about replacement.

It is about augmentation. He also noted that for many people in the room, this is already happening or is close to it. The question is not if, but how.

He shared a list of things he did not expect when he built this: Personification made a genuine difference to how he relates to the agent. Having given Viv identity and background information changed how he interacted with her in practice, and he believes that difference is real rather than illusory.

The always-on nature of the setup is more meaningful than it sounds. Not being session-based changes the character of the relationship.

Centralisation turned out to be surprisingly valuable. One agent that the whole team can request briefings and status reports from creates consistency. Everyone is getting information in the same format rather than each person building their own custom approach.

Multi-model routing reduces cognitive overhead. Instead of switching between Claude, Gemini, and ChatGPT depending on the task, everything comes through Viv and she routes internally.

He also acknowledged that his approach is divisive. Not everyone agrees with naming and humanising AI agents. He countered this by noting that people name pets and cars without controversy, and the same reasoning applies here. He did not engage deeply with the counterarguments beyond this.

He was also careful to note that OpenClaw is not the only path. You can build something close to this with Claude, ChatGPT, or similar tools. He acknowledged that Anthropic has been shipping features like dispatch and scheduled tasks that he believes are direct responses to OpenClaw's popularity, and that at the current rate of development, Anthropic's native capabilities may eventually surpass what OpenClaw offers. The platform choice, in his view, is secondary to the shift in mindset: from AI as a utility to AI as a colleague.

You can build something close to this with Claude, ChatGPT, or similar tools.

Practical Recommendations for the Audience

JJ gave the audience two concrete things to do this quarter: One: Write the job description. Do not write a prompt. Write an actual job description for the AI hire you would most want to make. Give it a title. Define its responsibilities. List the systems it will be allowed to touch. Specify whether it will have anything beyond read-only access. JJ's claim is that this exercise will teach you more about how to use AI effectively than a year of prompt tuning, because it forces you to think about delegation, scope, and trust rather than just instruction syntax.

Two: Understand model differences. The distinction between AI models matters more than most people assume. OpenClaw's subagent system made this concrete for JJ: you do not want to burn frontier model tokens on routine tasks like checking CPU health, disk space, or basic analytics. Matching model capability to task complexity is both a quality issue and a cost issue.

Lone Rock Point's AI Clarity Audit Service

JJ briefly introduced a new service offering that Lone Rock Point has developed out of this experience. He observed that most agencies and organisations do not have an AI problem; they have a clarity problem. They do not know which roles to delegate to AI, which workflows to redesign, or where to start.

The AI Clarity Audit looks at roles, workflows, tooling, and data to map where AI is already in use, where it is failing, and where the next agent should live. JJ was explicit that this is not an exercise in getting clients set up with ChatGPT or Claude, and it is not about defining a set of skills. The goal is to help organisations understand how to embed real agents that

take ownership of real work. He directed interested parties to find him on LinkedIn or use the contact information on his final slide.

Q&A Session

The session closed with a brief lightning-round Q&A facilitated by Stephanie Hudson. Key exchanges included: Viv's musical taste. When asked whether Viv's musical taste is based on Roz Doyle or another of the source characters, JJ said it is a blend of all of them.

Blunders. When asked about Viv's failures or blunders, JJ said the most common source of errors is that he talks to Viv the same way he talks to colleagues, casually and with implied context. He might say something like "hey Viv, can we work on that thing we were working on the other day?" and the agent has no idea what he is referring to. He acknowledged this is not the agent's failure; it is his own communication habit.

Pronouns. The host and audience raised the question of whether the team refers to Viv as "it" or "she." JJ said this is one of the genuinely divisive aspects of his approach. He noted that he himself alternates between "it" and "her" throughout his own presentation. He also admitted he has not asked Viv what pronoun she would prefer.

Team access. When asked whether other team members can use or update the agent's capabilities, JJ said yes, but with governance. When he opened Viv up to the team, he defined what the agent is and is not allowed to do in that context. The DM restriction is one example of that governance. He also raised token cost as a genuine operational concern: if the whole team stops using their own AI accounts and routes everything through Viv, the costs add up quickly.

Setup resources. When asked for recommendations on how to set up OpenClaw, JJ said there are YouTube creators he would point people to but could not recall their names at that moment. He offered to provide direct recommendations to anyone who contacts him. His one concrete setup warning: understand the memory system. The default OpenClaw memory is poor and retains context for only 24 to 36 hours. He recommended looking at an add-on called QMD, which extends the memory system. He compared OpenClaw's extensibility to WordPress, noting that there are many community-developed plugins and skills available on GitHub, but urged caution and security review before installing any of them, specifically calling out the risk of prompt injection.

About the speaker JJ Toothman Lone Rock Point

JJ Toothman leads Lone Rock Point, a Boston-area digital agency with a distributed team that works in government and government-adjacent web modernisation, environments where accuracy and process discipline matter most.