Adapt Progress Evolve · All work
In development · The Switchboard
Jarvis
The studio's calls come in here: one voice at the front, a fleet of specialised agents behind it.
Our AI chief of staff: a voice-first assistant backed by a fleet of specialised Claude agents. It's how the studio runs, not a product for sale, and its architecture is public on GitHub.
The plans on the wall
The architecture reference is public on GitHub: how the agents coordinate over a shared bus and one memory graph, including the numbers and the things that broke. The Jarvis system itself stays private.
The story
Jarvis is how we run the studio. You talk to it, and behind the voice sits a fleet of specialised Claude agents covering engineering, SEO, security, content, design review, research, email and calendar. They coordinate over a shared bus, which is agent-to-agent traffic, and they all read from and write to one persistent memory graph. Cortex Map was built to visualise that graph, and it's open source.
What do they do? They audit our sites, write content and quality-gate it, review designs with computer vision, watch the infrastructure and do research overnight. Anything that ships, spends or publishes goes through a human approval gate. Every time. The stack is TypeScript on Bun and Node, Postgres with pgvector, and Claude models tiered by cost, all on one server. The architecture reference is on GitHub, real numbers and real failures included.
Who it is for. Jarvis is for one person, and that person is the one running the studio. It isn't a product, and there's no waiting list. The problem it solves is plain: a small studio has more work than hands. Audits, content, security checks, email, calendar, research. Each one is a job on its own. So instead of hiring for every gap, you talk to an assistant and a fleet of agents picks up the pieces.
How it works. You speak, and the voice agent answers in real time, with 92 tool integrations behind it. Ask for something and the request is routed to whichever specialist owns that job: security, editorial, SEO, inbox, research, design and so on. The agents coordinate over a signed agent-to-agent bus, and every one of them reads from and writes to a single memory graph, currently 3,183 nodes and 3,924 edges. When the work is ready to ship, spend or publish, it stops and waits for a human to say yes.
What's different. Most assistants are one big model with a long prompt. Jarvis is a set of 22 separate services, each with its own repo, deploy and restart policy. A crash in the news agent can't take down voice. The bus has a Postgres failover, so the fleet keeps coordinating if its primary transport goes down. That's more moving parts than a single assistant needs. The trade is resilience and clear responsibilities, at the cost of more to run.
Where it stands. It's live and running, on one small Linux server. Voice latency over 847 real turns came in at a median of 3.0 seconds, with the slowest tenth at 6.3 or worse. That's good, not instant, and the published figures say so. Spend is capped hard per engine: $2 a night for insight mining, $15 a day for research, $20 a month for video. Forty-two repos sit on green CI. The numbers come from querying the live system, not from documentation.
What it deliberately doesn't do. It doesn't act on its own where it counts. Anything that ships, spends or publishes goes through the approval gate, every time, and no agent gets to skip it. It isn't multi-user, either. The system itself stays private and single-user, and only the architecture is public. The reference on GitHub includes the failures alongside the wins, because a fleet you can't trust to tell you what broke isn't worth running.
What it does
- Voice-first: you talk to it
- Specialised Claude agents for engineering, SEO, security, content and more
- Agents coordinate over a shared bus and one persistent memory graph
- Site audits, design review and overnight research
- Human approval gate on anything that ships, spends or publishes
- Architecture public on GitHub; Cortex Map is open source
More ventures
Cortex Map
Live
An open-source, MIT-licensed 3D knowledge-graph component for React, built on Three.js. Bring your own nodes and edges and it renders them as a glowing "war-table" memory map.
UKMWI
Paused
Tracked the UK's 100 most-searched companies, attention metrics, sentiment, momentum, sector performance, regional insights. Paused, the index is not updating for now.
ScriptGrain
Live
AI content that sounds like you, not a chatbot. ScriptGrain measures how you write across 45 reference points, builds a voice profile, and scores every draft against it.