Give your AI a body: your own Claude can read your record. healthsave mcp
prints one line that connects Claude Code or Claude Desktop to this server,
read-only.
Upgrade
healthsave update --to v1.4.0- No migration and no
.envchange on upgrade.healthsave mcpadds
MCP_READ_KEYto.envthe first time you run it. - The API image now also carries the MCP server
(/opt/healthsave-mcp/bin/healthsave-mcp, in its own environment; the
server's own packages are unchanged).
Added
- MCP read server (
packages/mcp, Give your AI a body). Eight
read-only tools: overview, today against your usual, sleep, activity, one
metric over time, findings, blood test results and Year in Health. Ask "how
did I sleep this year?" and it answers from your own record: nights, hours
asleep, bedtimes, stages, the nights with no data, month by month. Every
answer carries its units, the endpoints it read and plain limitation codes.
Two sources are never averaged. healthsave mcpprints the setup line for Claude Code, the config block for
Claude Desktop and a Docker line, with this server's address and a read key
filled in.healthsave mcp --checkstarts the server the way an AI client
does and calls one tool.healthsave mcp key --rotatereplaces the key.- A read key for the primary person. Until now a read key always belonged to a
participant. The new key reads your readings plus findings, correlations, lab
results and your time setting, so your AI client never needs the admin key.
A participant can never be bound to the primary person's record, even with
the same name as that key.
Privacy and safety
- The MCP server sends only
GETrequests, to the address you give it and
nowhere else; it does not follow redirects or use proxy settings. It hands
your AI per-day, per-night and per-month summaries, not raw readings; lab
results are listed one by one (at most 20 per marker). - No read key may call any route that writes. A test now fails the build if
one ever can. - What your AI reads goes to that AI's provider, like the rest of your
conversation. Your record stays on your hardware.
Not yet
- ChatGPT. It reaches MCP servers from OpenAI's cloud, so it would need your
Observatory open to the internet. - A read covers at most 5,000 readings per request; the tools say from which
day an answer is complete when a long period is cut.