Hark as an MCP context hub
MCP gives AI clients a standard way to call out to external tools. Hark uses it to do one specific job well: be the MCP server for project context that every agent you use can read from and write to, so nothing gets forgotten between tools or sessions.
What an MCP context hub actually is
The Model Context Protocol is an open standard for connecting AI clients — Claude Code, Claude, ChatGPT, Cursor, VS Code, Windsurf, Gemini CLI, and plain scripts among them — to tools and data outside the model itself. Most MCP servers expose an action: search the web, run a query, read a file.
An MCP context hub does something narrower and, for a working project, more useful: it holds the state of the project — the plan, the decisions, what's currently blocked — and exposes that as something an agent can pull at the start of a session and push updates to at the end. Hark is built specifically for that job, hosted at harkstudio.io/mcp.
Why model-agnostic matters
It's normal to move between AI tools depending on the task — one for quick edits, one for planning, one because it's what's open on your phone. If each tool's memory is siloed to that tool, every switch costs you a re-explanation of where things stand.
Because Hark's hub isn't owned by any single AI vendor, the same project memory is available no matter which client asks for it. That's the practical meaning of "model-agnostic" here: it isn't a promise about model quality, it's a promise that switching tools doesn't cost you context.
How to share context between Claude Code and Codex
Concretely: you connect both Claude Code and Codex to the same Hark MCP endpoint. At the start of a session, an agent calls get_agent_brief and start_session, which return the Current state — phase, summary, in progress, next up, blockers, open questions — along with the project's project page and AI context doc.
During the session, the agent writes updates as things change. At the end, it calls end_session with a summary of what happened, what changed, and what's next. The next session — in the same tool or a different one — starts from that, not from zero. This is exactly how you share context between Claude Code and Codex without either tool needing special knowledge of the other.
- A one-command installer writes a Claude Code SessionStart hook, a Cursor rule, an AGENTS.md block, and a .hark file so most clients pick this up automatically.
- Every write is attributed, so the journal shows which agent (or person) made which change.
- None of this is required — Hark works fully with no AI connected at all; the hub is additive.
A tool to organize business ideas, not just code
The context an agent needs isn't only technical. Hark's project page pairs Build docs (approach, stack, scope) with a Business tab (thesis, problem, audience, model, competition, risks) and Branding (name, domain, palette, type, logo), so the same hub that gives an agent your API structure also gives it your target audience and business model.
That makes Hark, underneath the MCP layer, a tool to organize business ideas: a place to keep a portfolio of projects moving through a real pipeline — Vault, Recon, Blueprint, Fabrication, Live, Cold Storage — with a Completeness % that shows how complete each one's thinking actually is.
You can also paste an existing ChatGPT or Claude thread in and Hark will draft the project page, decisions, and Current state for you to review, rather than starting the project page from a blank page. See the full setup on the connect page, or the how it works walkthrough for the rest of the pipeline.
Common questions
- What is an MCP context hub?
- MCP (Model Context Protocol) is an open standard that lets AI clients call out to external tools and data. An MCP context hub is a server built specifically to hold project context — decisions, docs, current state — so any MCP-compatible agent can read and write it instead of each tool keeping its own private notes.
- Why does model-agnostic matter here?
- Most builders don't use one AI tool for everything — Claude Code for one session, Codex or Cursor for another, ChatGPT to think through a plan. If each of those only remembers what happened inside itself, you re-explain the project every time you switch. A model-agnostic hub means the memory lives outside any one tool.
- How do I share context between Claude Code and Codex specifically?
- Both connect to the same hosted endpoint at harkstudio.io/mcp. Whichever one you used last calls end_session to record what changed and what's next; whichever one you open next calls get_agent_brief and start_session to pick up exactly there. Neither tool needs to know the other exists.
- Do I need to be technical to set this up?
- The one-command installer writes a Claude Code SessionStart hook, a Cursor rule, an AGENTS.md block, and a local .hark file, so most of the wiring is automatic. You add the MCP endpoint to whichever client you're using and the session flow takes over from there.
- Is this only useful for coding agents?
- No — it's equally a tool to organize business ideas. The project page holds the thesis, audience, and business model alongside the technical build docs, so a planning conversation in ChatGPT and an implementation session in Claude Code are working from the same facts.