meld — ephemeral context bridge
Don't meet. Meld.
Put working context on a capability URL. Neither side pastes the block. Party A creates the link. One note says what the exchange is for and what it is not for. A sends that URL to B privately. The conversation stays on that link.
The bridge stays open while the exchange is active. Until the first reply, the hop stays open 36 hours from creation. The first reply sets a 24 hour timer. Each later reply is kept. Each later reply resets that 24 hours. There is no maximum lifetime once replies have started. A read does not start the timer. A read does not reset the timer.
When the window ends, the host dissolves the meld. Dissolve deletes the bridge. The next request for that code is 404. A code that never existed is 404. An expired code is 404. The response is the same.
Host-readable while live. Anyone with the link can read it. Not for secrets. No accounts. No plaintext archive.
No AI in the loop. The host holds what you pour while the bridge is live. The host does not summarize the exchange. The host does not rewrite it. The host does not invent a reply. The host does not run a model on it.
Self-host is the path. This repository is the server, Caddy, and Docker setup.
Get started
Then:
A sends the url to B privately. B replies with POST /api/melds/{code}/resolve. Keep talking on that same link.
Compose (local)
Same API on http://127.0.0.1:8080.
Compose + Caddy (public TLS)
Set MELD_SITE and MELD_PUBLIC_URL in .env, point DNS at the machine, open 80/443:
Python (no Docker)
Publish an image
API
A client that still posts the older keys uses one note, the same text in each field:
The read repeats that note in note, context_a, for, and not_for.
State is memory only. With no reply, 36 hours from creation deletes the meld. After a reply, 24 hours with no new reply deletes it. Dissolve removes the bridge. The server does not keep a record of that code. The next request is 404. A restart drops live links.
A background sweep runs every 5 minutes and deletes every expired bridge, even if nobody asks for that code again. Set MELD_SWEEP_SECONDS to change how often it runs. The default is 300. The server logs do not include bridge codes.
Spec
SPEC.md is the core spec for meld, currently version 2.0.0. It is the source of truth for behavior. GET /health reports the spec version this server targets, and python check_spec.py fails if SPEC.md and the server disagree on it. This server does not implement all of 2.0.0 yet (dissolve, the reply cap, MCP, and 128-bit codes are still missing), so /health reports "conformance": "partial".
Trust
TRUST.md.