Skip to content

HarnessRouter Community Edition: the self-hosted, Apache-2.0 edition of the unified interface for agent harnesses. Run Codex, Claude Code, Hermes, PI, DSH, and more through one API, with sessions, streaming, files, cancellation, and failure handling. Implements the Unified Harness Protocol (UHP), an open standard. Your keys, your infrastructure.

fix: a media-only integration could never be saved, and blocked every other one

richard-epsilladae5dbe3c041
diff --git a/docker/entrypoint.sh b/docker/entrypoint.sh
index 76a3d572..22ad09d9 100644
--- a/docker/entrypoint.sh
+++ b/docker/entrypoint.sh
@@ -103,6 +103,15 @@ elif [ "$HR_AUTH_PASSWORD" = "harnessrouter" ]; then
echo "[harnessrouter] WARNING: using the DEFAULT password. Set HR_AUTH_PASSWORD, or change it from the profile page, before exposing this instance."
fi
export PORT="${PORT:-3000}"
+# Next binds to $HOSTNAME, and Docker sets that to the container id, which resolves to ONE of the
+# container's addresses. That is fine with a single network and silently fatal with two: connect
+# this container to a user-defined network — which is exactly what the README tells you to do to
+# reach a database — and after the next restart Next comes up on that network's address while the
+# published port still forwards to the bridge one. Nothing is listening where the port lands, so
+# the console answers 502 while the container reports healthy and the log says "Ready".
+# Measured on the test box 2026-08-16: LISTEN 172.18.0.5:3000, published 127.0.0.1:3000 -> the
+# bridge ip. Binding every interface is the only answer that survives a second network.
+export HOSTNAME=0.0.0.0
# ── agent CLIs: installed here, not shipped in the image ──────────────────────
# Claude Code is distributed under Anthropic's own terms and hermes-agent declares no license,
diff --git a/gateway/app.py b/gateway/app.py
index b8f58329..be29d1e6 100644
--- a/gateway/app.py
+++ b/gateway/app.py
@@ -3469,7 +3469,13 @@ async def admin_integrations_put(body: IntegrationsBody, request: Request) -> di
for i in body.integrations:
name = str(i.get("name") or "").strip()
provider = str(i.get("provider") or "").lower()
- if not name or provider not in {p for p, _ in _INTEGRATION_WIRING}:
+ # Against the CATALOG, which is the list of providers this console offers, and not against
+ # _INTEGRATION_WIRING, which answers a different question: which agent backends can be
+ # driven by one. A media-only provider has no backend wiring by definition, so validating
+ # here against the wiring made ElevenLabs unsaveable — and because the console PUTs the
+ # WHOLE document, one such row already in the vault failed every subsequent write and no
+ # integration of any kind could be added to that instance again.
+ if not name or provider not in _PROVIDER_CATALOG:
raise HTTPException(400, f"integration needs a name and a known provider (got '{provider}')")
cfg = {k: v for k, v in (i.get("config") or {}).items() if v not in (None, "")}
# The client never sees secrets (sentinel) — carry stored values through unchanged edits.