We automate the work your team keeps putting off.
A boutique AI agency based in Manchester, UK. AI SDKs, MCPs, developer tooling - or building websites or web applications at a rapid pace.
Prefrontal
PhonePe
The Agent-Ready API sprint.
Your HTTP API already works for people and for standard integrations. Now your customers want to point LLM agents at it, and the agents hit a wall: no OpenAPI spec, no SDK, no MCP server, no docs a model can read. This sprint fixes that. Four artifacts, generated from your existing API, shipped in two weeks.
-
An OpenAPI spec that matches reality
Cleaned up if you have one, generated from the code if you don't. Every endpoint, parameter, and error shape, accurate to what the server actually does.
-
TypeScript and Python SDKs
Generated from the spec and published to npm and PyPI. A customer integration starts with an import.
-
A published MCP server
Ready for your customers to plug into Claude, Cursor, or their own agent stack the day it ships.
-
Documentation agents can read
llms.txt, tool descriptions, and worked usage examples, written so a model can discover what your API does and call it correctly on the first try.
Scope it free. Ship it in two weeks.
-
01 / Scope
A free 30-minute call
We go through your API together. You leave with a written scope for the sprint and a straight answer on whether it is worth running.
-
02 / Build
Two weeks, £10-12K
Spec, SDKs, MCP server, and agent docs, generated from your codebase and delivered as working, published artifacts. Fixed scope, fixed price.
-
03 / Handover
Yours, documented, running
Everything lands in your repos with the generation pipeline included, so the artifacts stay current as your API changes. Follow-on work beyond the fixed scope is £750 a day.
You might be a fit if…
- You've raised a Series A or B, and your public HTTP API is how customers integrate with you.
- Those customers have started asking to point Claude, ChatGPT, or their own agents at it, and there's nothing to point them at.
- Your OpenAPI spec is missing, or drifted so far from the code that nobody trusts it.
- An engineer picked up "the MCP thing" a quarter ago, between other priorities, and it's still a branch.
- You want one engineer accountable for the whole thing, from spec to shipped MCP server, with everything written down.
Where this has already shipped.
- Meta. Akash came up through Meta's rotational engineering programme before founding Jōshi Systems.
- DeepL. Built the public DeepL MCP server, now at 10k+ monthly users, plus 15 internal MCP connectors, and shipped the DeepL Voice API end to end.
- Prefrontal AI. The exact build described above, delivered as a three-week sprint in June 2026. They kept him on, and the contract is still running.
I've spent my career shipping AI products inside serious engineering organisations, most recently at Meta and a Cambridge AI lab. I started Jōshi Systems because the most useful AI work isn't happening at hyped-up startups — it's waiting inside operations teams at real companies, and it mostly needs someone to just build the thing.
Akash Joshi / Founder
Bring the API. Leave with a scope.
A free 30-minute call. We go through your endpoints and tell you exactly what the sprint would ship for you.
Book a free scoping call →