
Keeper Secrets Manager
io.github.WYRE-AIPublished by GitHub account WYRE-AIv1.0.4Updated Oct 24, 2026
Read-only multitenant Streamable HTTP bridge over Keeper Secrets Manager's MCP server.
The publisher's own server: Secrets.RealEstate
Overview
Keeper Secrets Manager's local MCP server; no sign-in required; STDIO.
No AI overview yet: these facts come from the registry entry and the server itself.
Installation
In SourceWeft
- Open Keeper Secrets Manager in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
keeper-mcp
Multitenant Streamable HTTP bridge over Keeper-Security/keeper-mcp-golang-docker — Keeper's official Secrets Manager MCP server — built so the WYRE conduit gateway can forward per-tenant Keeper credentials as HTTP headers.
Upstream pin:
v2.5.0(MIT), built from source. See Bumping the upstream pin.v1 is READ-ONLY. See Read-only in v1 — for this vendor that is a security boundary, not a scoping decision.
Why
Keeper's server is stdio-only. There is no HTTP or SSE listener anywhere in its source — the EXPOSE 8080 in its own Dockerfile is vestigial, and its docker-entrypoint.sh is unreachable because the published image is FROM scratch with no shell. It reads its credential from KSM_CONFIG_BASE64 at process start, which makes it single-tenant per process.
Our gateway is multi-tenant: every request carries the calling org's credentials as HTTP headers, and the vendor container has to translate those headers into something the upstream understands.
Because the upstream has no HTTP mode to proxy to, this bridge holds an MCP client session over stdio to each tenant child and re-serves it over Streamable HTTP:
- Listens on
:8080withPOST /mcpandGET /health. Refuses to start unlessCONDUIT_S2S_SECRETis a real secret. - 401-gates every
/mcprequest onX-Gateway-S2S, then onX-Keeper-Config-Base64— including validating that it decodes to a usable KSM config. Missing/invalid credentials never fall through to environment credentials; that would be a cross-tenant leak. This bridge is gateway mode only. - Lazily spawns one
ksm-mcp serve --batchchild per credential (keyed by a hash), with the tenant'sKSM_CONFIG_BASE64set, and connects an MCP client to it over stdio. - Serves both protocol eras on
/mcpvia the v2 SDK'screateMcpHandler(factory, { legacy: 'stateless' })— 2025-erainitialize-handshake clients (the conduit gateway today) and modern 2026-07-28 envelope clients. - Filters the child's surface through a read-only allowlist in both directions.
- Evicts idle children after 15 minutes (
IDLE_EVICT_MS).
Tool names pass through unchanged.
Read-only in v1 — deliberate, and load-bearing
Boundary 1 (the strong one): a read-only KSM application
Provision the Keeper application with non-editable shares. A KSM application's
shared-folder grant is read-only unless explicitly made editable
(secrets-manager share add --app <APP> --secret <FOLDER_UID>; the Vault's Application
Access panel shows Read-Only vs Editable). With a non-editable grant, every write
fails at Keeper's own server — regardless of this bridge, its allowlist, batch mode, or
prompt injection.
This is a provisioning precondition of the credential contract, not a nice-to-have.
The bridge cannot verify it: the Keeper SDK carries IsEditable per record, but
ksm-mcp never reads it, so nothing here can detect an over-granted application. Scope
the application to the narrowest folder set the use case needs, read-only, and treat that
as the real control.
Boundary 2: this bridge's allowlist
Keeper's upstream has no read-only mode — verified against the v2.5.0 pin: serve has six
flags and none restrict the surface, getAvailableTools() returns a hardcoded 19-element
slice that consults no config, and Confirmation.DefaultDeny is hardcoded false with no
plumbing. Nothing upstream can un-register a tool.
On batch mode. Keeper's own Docker documentation tells you to set
KSM_MCP_BATCH_MODE=true, and this bridge does. What that actually changes:
Leaving it off would not buy a confirmation boundary here, because that two-phase flow is
unreachable through this bridge by construction: it is driven by an MCP prompt, and the
bridge never declares the prompts capability — and the tool that completes it,
ksm_execute_confirmed_action, is blocked below. So "off" yields stubs no client here can
satisfy, not safety. We force it on and put the boundary in src/tools.ts instead.
What the allowlist uniquely buys, which a read-only grant does not: blocking
get_all_secrets_unmasked and ksm_execute_confirmed_action. Both are reads-or-worse
that a read-only grant happily permits.
Served (9)
Conduit classifies credential reads as admin (src/access/tool-classification.ts), which
outranks write. A Keeper read is a credential read by definition.
Blocked (10)
create_secret, update_secret, delete_secret, create_folder, delete_folder,
upload_file, ksm_execute_confirmed_action, get_all_secrets_unmasked,
get_record_type_schema, download_file
Four are worth explaining:
ksm_execute_confirmed_action— not a peer of the others but the keystone. It takesoriginal_tool_nameplus auser_decisionboolean the caller supplies, with no nonce and no correlation to any prompt, then dispatches straight into the confirmed executors. Allowing it would collapse every other rule here into one name. Blocked permanently.get_all_secrets_unmasked— one call that returns every secret in the application's scope, unmasked, into the model's context. Blocked permanently.get_record_type_schema— non-functional upstream at v2.5.0:LoadRecordTemplates()has no callers and the package has noinit(), so every call returns "record templates not loaded".download_file— cannot work over MCP. Upstream'sDownloadFile(uid, fileUID, savePath)returns only anerrorand writes bytes to a server-side path; the handler returns{uid, file_uid, path, message}, never the file. Honouringsave_pathwould let a tenant write attacker-chosen paths into a container filesystem shared with every other tenant's child, and removing it leaves nowhere to write.
Argument policy: allowlist, not blocklist
Each served tool declares the exact arguments it accepts (allowArgs in src/tools.ts),
enforced on both the advertised inputSchema and the inbound tools/call. A client
working from a stale or ignored schema therefore cannot reach a write path by passing an
argument anyway.
The default matters more than the current entries: when an upstream bump adds a parameter
to an already-served tool, it fails closed — it vanishes from the surface until
someone reviews it into the table — rather than passing through silently. Today the only
removals are generate_password's save_to_secret and folder_uid, which would turn a
local utility into a record create.
Enabling writes is a reviewed, versioned change to src/tools.ts plus the matching
VENDOR_TOOL_CONFIG rows in conduit. Not a config flip, not an env var.
Credential contract
The gateway forwards this header on every /mcp request. conduit's vendor-config must match exactly.
The value is the base64 device configuration from Keeper Vault → Secrets Manager → your application → Devices → Add Device (it starts ewog...).
Precondition: the application this device belongs to must hold read-only (non-editable) shares. See Boundary 1 — it is the strongest control in the system and the bridge cannot verify it for you.
Validity rule: standard base64 that decodes to a flat JSON object of string values containing clientId, privateKey and appKey. hostname is optional — verified against upstream internal/ksm/client.go InitializeWithConfig, which requires exactly those three. Anything else → HTTP 401 with a JSON-RPC error body that never echoes the supplied value back.
The bridge also forces on every child:
KSM_MCP_PROFILE is deliberately not set: with KSM_CONFIG_BASE64 present, upstream builds an in-memory profile and never touches the on-disk profile store, which is what keeps tenants from sharing state through the filesystem.
Configuration
Copy .env.example. Leave CONDUIT_S2S_SECRET empty in the example; put a real value only in an untracked .env or the process environment.
HTTP authentication
With CONDUIT_S2S_SECRET set, every /mcp request must include X-Gateway-S2S. The check is verifyS2sHeader in src/s2s-verify.ts: the header value is t=<unix seconds>,v1=<hex>, where the hex is HMAC-SHA256 of the literal string t=<unix seconds> keyed with CONDUIT_S2S_SECRET (64 lowercase hex characters). The timestamp must be within 300 seconds of the server clock. A missing or invalid header is HTTP 401, before the Keeper credential header is read.
Gateway mode is the only mode. /mcp also requires X-Keeper-Config-Base64. That header is the only credential source for the request. KSM_CONFIG_BASE64 in the process environment is not a fallback.
/health and /healthz stay unauthenticated and do not read credentials.
Docker
The image listens on 0.0.0.0 inside the container (MCP_HTTP_HOST). Publish the port on loopback and pass a real S2S secret. Keeper credentials arrive on each request as headers; they are not read from the environment.
Compose refuses to start when CONDUIT_S2S_SECRET is unset. The gateway service does not set KSM_CONFIG_BASE64 or env_file, and publishes 127.0.0.1:8080:8080:
Local HTTP without a secret is development-only. It binds 127.0.0.1 and logs a warning. A non-loopback MCP_HTTP_HOST (including 0.0.0.0) is ignored:
Endpoints
POST /mcp— Streamable HTTP MCP, dual-era. Requires a validX-Gateway-S2Sheader andX-Keeper-Config-Base64.GET /healthandGET /healthz— unauthenticated liveness. Reportsmode, served tool count, and live tenant count. Never reads credentials and never spawns a child.
Development
HTTP refuses to start until CONDUIT_S2S_SECRET is set. For a local listener without a secret:
The test suite covers the allowlist policy, the credential contract, and the 401 gate over the real HTTP stack. None of it needs a Keeper account or a running child — the pool is pointed at a nonexistent binary so any accidental spawn fails loudly.
Bumping the upstream pin
- Change
KSM_MCP_REFin theDockerfile. - Rebuild.
scripts/check-upstream-tools.mjsruns in the Docker build, re-derives the tool surface from the pinned Go source, and fails the build on any drift — a new tool, a renamed one, or a new argument on an already-served tool. This is enforcement, not a checklist item. - Triage whatever it reports into
src/tools.tsandsrc/__tests__/fixtures.ts. New tools are default-denied by the allowlist, but they still need an explicit decision recorded. - Re-run the build; both build-time smoke tests and the drift check must pass.
- Cut a release, then re-pin the digest in conduit's
azure/vendor-fleet.conduit-prod.bicepparam.
Never track main, and never pull keeper/keeper-mcp-server:latest — the pin is a reviewed git tag we build ourselves.
License
Apache-2.0 (this bridge). The upstream Keeper server is MIT and is built from source at image build time.
Source: README.md at commit 1b22549
Tools
0Version history
3- v1.0.4LatestOct 20, 2026
- v1.0.3Sep 22, 2026
- v1.0.2Sep 21, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts
Other Secrets integrations
Secrets.RealEstate
realestate.secretsOfficial publisher
A remote MCP endpoint giving assistants sourced French property intelligence: sold prices, energy ratings, risk, crime and school data for addresses.

Bitwarden Public API and Secrets Manager (unofficial)
Published by GitHub account mguttmann
Unofficial local MCP server that lets an assistant manage Bitwarden organization members, groups, collections, policies and Secrets Manager projects and…

LLM Data Guard: PII & Secrets Redaction
Published by GitHub account fairquery
Detect and redact PII and secrets (API keys, tokens) in text, HTML or PDF before it reaches an LLM.

Keeper Secrets Manager
Published by GitHub account wyre-technology
Read-only, multitenant HTTP bridge that re-serves Keeper Secrets Manager's MCP server so an assistant can list, search and read vault secrets per tenant.

Hive Secrets
Published by GitHub account srotzin
A remote MCP server that stores and retrieves encrypted agent credentials in per-agent namespaces, with paid reads and writes settled in USDC on Base.