Signet
A provider-neutral MCP human approval gateway that requires explicit confirmation before executing write operations, supporting adapters like Fastmail and WhatsApp.
README
Signet
Signet is a provider-neutral MCP human approval gateway. A configured write call
returns an honest pending_approval result only after the exact executable
payload, expiry, origin namespace, and byte-identical acknowledgement are durable.
That acknowledgement never claims the external action succeeded. The downstream
mutation remains unsent until a fresh human confirmation authorizes the frozen
request version.
The first reviewed adapters cover Fastmail email and an owned wacli WhatsApp
wrapper. The core is generic: exact MCP schemas are mirrored behind four policy
modes, immutable payloads are encrypted, approval transitions are persisted in
SQLite, dispatch is fenced, ambiguous delivery enters bounded reconciliation, and
the authenticated web app presents the private review queue.
Safety status
This repository is in no-live implementation mode. Tests use explicit fake:*
identities and fake downstreams. No repository command enrolls a passkey or TOTP,
reads a live provider credential, sends a real message, changes an existing Hermes
profile, installs a launchd job, changes Tailscale Serve, or performs cutover. The
documented helpers modify only a newly created blank fake or downstream-disabled
profile and may ingest its Signet caller token through standard input.
The files under deploy/ are inert review templates. Their placeholders prevent
installation without review. The installed signet deployment commands provide a
runnable downstream-disabled staging assembly: it has no provider transport,
credential resolver, dispatch worker, or downstream MCP alias. Its authenticated
approvals tools all return deployment_disabled. This is not a live deployment
or a substitute for deferred human setup and cutover. signet.operations consumes
local, bounded fixtures only; it has no discovery network client or host scanner.
Guarantees
- Unknown tools resolve to
deny. A tool is exposed only after exact schema capture, policy configuration, and digest review. approvaltools make zero downstream calls before approval and return the normative pending shape inspec/fixtures/gateway-pending-result.json.- A fresh TOTP proof or WebAuthn assertion is bound to one action, request, immutable version, and payload hash, then consumed transactionally.
- The MCP TOTP path can approve a normal caller-owned request, but cannot deny, edit, retry, manage credentials, or approve a policy change.
- Dispatch crosses a durable fenced boundary before network I/O. A possible
post-dispatch crash becomes
outcome_unknown, never a blind retry. - Production retains exhausted unknown content indefinitely. The demo-only redaction drill is marker-guarded fake functionality, preserves "may have sent", and is not production human authorization.
- Push messages contain category and count information only. The authenticated queue remains authoritative if push delivery fails.
- Provider credentials are references such as
keychain://Signet/fastmail, not values accepted by normal configuration models. - The MCP listener is loopback-only. The separately bound web app supplies its own login, sessions, CSRF validation, action confirmation, and security headers.
These controls protect managed MCP routes. They do not prevent a malicious process
running as the same operating-system user from reading that user's files, memory,
or Keychain items, and they cannot govern direct provider scripts, native adapters,
browser sessions, webhooks, or other paths that bypass Signet. See
docs/security-model.md.
Development
Signet requires Python 3.12 and uses uv. The repository pins uv-managed Python
3.12.13 because its bundled SQLite satisfies Signet's 3.51.3 safety floor.
uv python install 3.12.13
uv sync --frozen
uv run pytest -q
uv run ruff check .
uv run mypy
The generic reviewed local stdio boundary currently requires Linux with
/proc/self/fd. macOS remains supported for the downstream-disabled launchd demo
and separately reviewed HTTPS downstreams, but local process activation fails
closed with process_boundary_platform_unsupported. The sole reviewed wacli
fixture pins a macOS Homebrew artifact, so wacli activation is blocked on every
host until either a Linux artifact is reviewed or a secure native macOS descriptor
boundary is implemented and characterized.
The generic package entry point serves only an explicitly supplied application factory. After creating the disabled state below, its two shipped factories can be run explicitly with:
export SIGNET_DISABLED_CONFIG="$HOME/.hermes/services/signet/config/disabled.json"
uv run signet serve-mcp --factory signet.deployment:create_mcp_app \
--host 127.0.0.1 --port 8789
uv run signet serve-web --factory signet.deployment:create_web_app \
--host 127.0.0.1 --port 8790
SIGNET_DISABLED_CONFIG is the absolute non-secret config path, not configuration
JSON or a credential. The dedicated signet deployment serve-* commands below are
preferred because they use the verified listener settings from that file. The MCP
command rejects a non-loopback numeric host. Do not point an ad hoc factory at live
credentials.
Downstream-disabled deployment staging
Create private, persistent staging state without enrolling a human credential or creating a downstream client:
export SIGNET_DISABLED_PROFILE=signet-disabled
install -d -m 0700 "$HOME/.hermes/services/signet/config"
uv run signet deployment init \
--config "$HOME/.hermes/services/signet/config/disabled.json" \
--data-dir "$HOME/.hermes/services/signet/data" \
--namespace "profile:$SIGNET_DISABLED_PROFILE"
uv run signet deployment validate \
--config "$HOME/.hermes/services/signet/config/disabled.json"
uv run signet deployment serve-mcp \
--config "$HOME/.hermes/services/signet/config/disabled.json"
One initialized disabled state supports exactly that one dedicated Hermes profile;
the CLI does not add principals to an existing config. Follow the tested
deploy/hermes/README.md
sequence to create the blank profile and stream token issue directly into the
checked-in atomic configurator. The raw token is accepted only on stdin and is never
written to YAML or output by the helper. Do not paste it into an argument, shell
history, log, chat, or documentation. token list returns metadata only. token revoke --token-id=TOKEN_ID takes effect on the next authentication check. token rotate --token-id=TOKEN_ID stages and
prints a linked replacement while deliberately leaving the old token valid; install,
reload, and test the replacement before explicitly revoking the old token.
The optional init human-auth context flags validate only the exact HTTPS origin,
RP ID, and user ID. deployment auth-status reads counts, not credential material.
Neither command enrolls anything. A passkey requires a real browser/authenticator
ceremony at the final HTTPS origin and cannot be created by an offline CLI. See
docs/deployment.md.
Fake-only operator path
docs/operator-runbook.md is the start-to-finish path
for a disposable local demo, Hermes profile wiring, verification, troubleshooting,
backup/restore drills, and rollback. The demo uses explicit fake identities and
network-disabled providers; it is not evidence of passkey/TOTP enrollment, live
schema review, provider readiness, or cutover authorization.
From the repository root, the minimal fake-only path is:
export SIGNET_DEMO_DIR="$PWD/var/operator-demo"
test ! -e "$SIGNET_DEMO_DIR"
uv run signet demo init --data-dir "$SIGNET_DEMO_DIR"
uv run signet demo smoke --data-dir "$SIGNET_DEMO_DIR"
uv run signet demo serve --data-dir "$SIGNET_DEMO_DIR"
demo init refuses every existing path, smoke is offline unless --live is explicit,
and serve binds both demo apps to numeric loopback. The generic serve-* factory
interface remains deployment-owned. Hermes templates stay inert; the runbook uses a
new blank profile and a validated structured merge instead of editing an existing
profile.
Offline onboarding
Operational helpers are available without adding another console entry point:
uv run python -m signet.operations --help
They normalize a previously captured local tools/list fixture, add advisory
read/write hints, generate an all-deny policy, create and verify fake-adapter test
inputs, evaluate a caller-supplied names-and-locations-only bypass inventory, and
produce a fail-closed cutover readiness report. Output files are created once with
mode 0600; existing files are never overwritten.
The readiness report is advisory: it always keeps ready and
authorizes_live_changes false, even when its supplied evidence packet is complete.
Repository map
spec/contains executable policy, provider-input, pending-result, and gateway tool schema fixtures.src/signet/contains canonicalization, encryption, persistence, authentication, MCP mirroring, gateway, adapters, web UI, notifications, backup, and operations.tests/contains contract, adversarial, durability, authentication, adapter, runtime, web, backup, and offline operations coverage.docs/contains the MCP tool reference, security model, deployment guide, and policy/onboarding guide.deploy/contains secret-free launchd, Homepage, Tailscale, Hermes, and readiness staging material. It changes nothing by itself.
Documentation
- MCP approval tools
- Security model
- Policy and adapter onboarding
- Deployment, backup, restore, and rollback
- No-live operator and Hermes runbook
The implementation contract and deferred human-only ceremony are recorded in
2026-07-14-signet-approval-gateway-plan.md.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。