terse sits between an MCP client and its downstream servers, wrapping and rewriting
tool traffic. Its trust model: the client and the downstream servers it wraps are
trusted (they're config the user chose to run); a downstream server's tool output
is not — it can come from a hostile or compromised source (e.g. a URL fetched from a
repo-committed .mcp.json, per transport.py), so terse treats decoding and diffing
of tool results as adversarial input, not just plumbing. Config files terse writes or
touches (.terse-peers*.json, .claude.json stashes, backups) can carry API keys and
other credentials inherited from wrapped servers' env blocks; these are written
0600 and excluded from lossy transforms, but that does not stop a git add -A in
project scope — see USAGE.md's guidance on .gitignore.
HTTP downstreams and environment proxies: an HTTP downstream honors HTTP_PROXY /
HTTPS_PROXY / NO_PROXY like any urllib client. terse's URL guards (scheme allow-list,
metadata-address refusal, cleartext-credential refusal) judge the downstream URL, not the
proxy: with an http:// loopback downstream and a configured HTTP_PROXY, the request —
credential headers included — travels through that proxy. Set NO_PROXY=localhost,127.0.0.1
(or unset the proxy) for local gateways that carry a key. Through a proxy, terse's
connect-time check sees the proxy's address, not the target's, so a host name that resolves
to a cloud-metadata address is only refused when there is no proxy in the way.
Out of scope: vulnerabilities in wrapped downstream MCP servers themselves (report to their maintainers), and threats requiring local filesystem access an attacker already has (terse does not attempt to defend against a compromised host).
Open a GitHub security advisory for this repository, or file a regular issue if the report doesn't need to stay private. Please include a minimal repro (which command, which config shape) — terse has no telemetry, so reports are the only signal maintainers get.