● early access · relay.codeclowns.com
Relay
A memory your AI tools share.
Captures the decisions, constraints, open questions and rejected approaches from your work with AI, as structure rather than a pile of text.
chat · Monday“No new dependencies in the billing package.”
relayconstraintbilling · no new depssource: chat, Mon 10:42
code agent · ThursdayUses the existing HTTP client instead of adding one.why? ↗ constraint from Monday
How Relay works
- 01Work as usualChat, plan and code in whichever AI tools you already use.
- 02CaptureDecisions, constraints, open questions and rejected approaches are pulled out as structure.
- 03ProvenanceEach one keeps a link to where it came from, and you can correct it.
- 04RecallThe next tool asks Relay over MCP and gets only what applies to the task.
- 05It just knowsA fresh session obeys a constraint it was never told, with a chip showing why.
Under the hood
in
Web chats, via the browser extension
Coding agents, via MCP
The desktop app
The local daemon
Provenance storeevent-sourced, every fact traceable
On-device modelsbreak work into typed units
MCP brokerone interface for every tool and agent
Lean recallonly what the next task needs
out
Claude and other chat tools
Claude Code, Cursor and other agents
Team sync, later
Today: relay-core runs locally with an MCP server, a CLI, a local HTTP API and SQLite. The browser extension is in development.
how it uses AI
- A local context layer that runs on your machine, with an MCP server that Claude and coding agents can read from and write to.
- On-device models that turn conversations into typed units: decisions, constraints, arguments.
- Recall that returns only what the next task needs, so context stays lean.
for
Developers and teams who work across several AI tools and are tired of re-explaining their project.
the line it holdsYour context is yours: stored locally, portable between tools, never deposited in one vendor’s memory.
Where Relay is going
- NowSolo developersNever re-explain your project to a new agent.
- NextSmall teamsShare a project’s decision log; onboarding is pointing someone at Relay.
- LaterCompany memoryInstitutional memory with admin, single sign-on and audit.
- VisionShared memory for agentsAgents read and write context through Relay, which keeps provenance and keeps them apart.
Structure, not a pile of textDecisions and constraints as typed units with sources are smaller, cheaper and more reliable to recall than raw history.
Yours, not a vendor’sContext lives on your machine and moves between tools, instead of being locked inside one product’s memory.
Models built for the jobSmall on-device models tuned for turning conversations into decisions, so capture is private and costs nothing per call.