AgentWire is a 100% free and open-source observability, governance, and security gateway specifically designed for modern AI Agents, Large Language Models (LLMs), and Model Context Protocol (MCP) servers. Just as Wireshark inspects network packets and Cloudflare provides a protective edge layer, AgentWire sits between your users, agents, and external services to monitor, inspect, analyze, secure, replay, and optimize every single AI interaction.
OpenTelemetry + Wireshark + Cloudflare for AI Agents. Monitor, inspect, secure, replay, and optimize all traffic between Users, AI Agents, LLMs, and MCP Servers.
ดาว
4
เครื่องมือ
—
อัปไทม์
—
[!NOTE]
100% Open Source Commitment: AgentWire is built from the ground up to be 100% free and open source under the Apache 2.0 License. Multi-tenancy, RBAC, SSO (OIDC + SAML), the security scanner, and the audit log are included directly in the open-source repository with zero commercial paywalls, open-core restrictions, or proprietary tiers. Custom (user-configurable) security rules and the distributed storage architecture described below are still on the roadmap.
✨ Features
Everything below is real, working code with an automated test covering it (57 tests, tests/AgentWire.Tests) — not aspirational copy. See docs/roadmap.md for what's still ahead.
🕵️♂️ AI Packet Inspector: Ingest traces via POST /v1/traces (per-organization API key), list/inspect them via GET /v1/packets.
⏪ Replay Engine: POST /v1/packets/{id}/replay re-sends a stored prompt to any OpenAI-chat-completions-compatible endpoint (OpenAI itself, or a local Ollama), with optional model/temperature override. Returns a clear 422 (not a fake success) if no provider is configured.
🛡️ Security Scanner: Rule-based (regex, not ML — see docs/roadmap.md for the "custom rules engine" item this isn't) detection of prompt-injection phrasing and PII (email, phone, Luhn-validated credit cards, SSN) on every ingested and replayed packet. Findings are masked before storage — the finding table never holds raw sensitive text.
🏢 Multi-Tenancy: Every packet, finding, and audit entry is scoped by OrganizationId and enforced per-query. One caveat: this build only provisions a single organization per self-hosted instance (POST /v1/setup is a one-time bootstrap) — there's no multi-org provisioning endpoint yet.
🔐 RBAC: Admin/Member roles enforced via JWT claims ([Authorize(Roles=...)]) on every admin-only route (users, API keys, audit log).
🪪 SSO — OIDC and SAML 2.0: Both are real integrations, not stubs. OIDC uses standard ASP.NET Core OpenID Connect middleware. SAML uses ITfoxtec.Identity.Saml2 for SP-side metadata/login/ACS, with a self-signed SP certificate generated on first run. Honest scope boundary: both are validated by automated tests that either exercise the real middleware (OIDC) or hand-sign a test SAML assertion against a locally-generated fake IdP (SAML) — proving the SP-side pipeline works, not interop with a specific real-world IdP (Okta, Entra ID, Keycloak). One upstream IdP per self-hosted instance, not per-organization.
📜 Immutable Audit Log: GET /v1/audit-log (Admin-only) — every login, API-key change, role change, and replay attempt is recorded. Immutability is enforced by having no PUT/PATCH/DELETE route under /v1/audit-log, not a soft-delete flag.
💸 Cost Intelligence: GET /v1/analytics/costs — cost rollups by model, scoped to your organization.
Needs an OpenAI-compatible provider configured (see Replay:BaseUrl/Replay:ApiKey below) — without one, this returns a 422 explaining exactly that, rather than faking a response:
Both are disabled by default. Set in appsettings.json or environment variables:
OIDC: Oidc:Enabled=true, Oidc:Authority, Oidc:ClientId, Oidc:ClientSecret → GET /v1/auth/oidc/login.
SAML: Saml:IdpMetadataUrl or Saml:IdpMetadataXmlPath → GET /saml/login, ACS at POST /saml/acs, SP metadata always served at GET /saml/metadata (hand this to your IdP). One upstream IdP per instance.
Both funnel into the same local-user model: the IdP authenticates identity only, AgentWire finds-or-provisions a local user (always Member, never auto-Admin) and issues its own JWT for everything else.
7. Optional — dashboard scaffold
cd src/Client/dashboard
npm install
npm run dev
Open http://localhost:3000 — this is still an unmodified create-next-app scaffold with no real pages wired to the API yet.
8. Optional — local infrastructure stack
deploy/docker-compose.yml starts Postgres, Redis, RabbitMQ, and ClickHouse — none of these are wired to any code path yet (see Architecture below). The running API uses SQLite only.
Running the tests
dotnet test tests/AgentWire.Tests
🏗️ Architecture
What's actually running today: a single ASP.NET Core 10 API (src/AgentWire.Presentation) backed by SQLite/EF Core, with JWT-claims RBAC, ApiKey auth for ingestion, OIDC + SAML SSO, a rule-based security scanner, and a replay engine calling out to any OpenAI-compatible provider. deploy/docker-compose.yml's Postgres/Redis/RabbitMQ/ClickHouse services exist but aren't wired to any code path.
The distributed, polyglot architecture (YARP gateway, RabbitMQ ingestion pipeline, ClickHouse for packets, Postgres for metadata, Redis caching) described in the original design doc is still aspirational — kept as the scale-out direction, not what's deployed.
AI AgentsLLM ObservabilityAI SecurityPrompt InjectionMCP ServersModel Context ProtocolAI GatewayLLM AnalyticsAI Traffic MonitorGenerative AIAgentic AIAI InfrastructureMachine LearningLLMOpsAIOpsOpenTelemetry for AIWireshark for AICloudflare for AIOpen Source AIFOSS AI
📝 License
This project is 100% Free & Open Source and licensed under the Apache 2.0 License - see the LICENSE file for details.
A
AgentWire
qmmughal·4 ดาว·
OpenTelemetry + Wireshark + Cloudflare for AI Agents. Monitor, inspect, secure, replay, and optimize all traffic between Users, AI Agents, LLMs, and MCP Servers.
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
AgentWire ต้องใช้ API key ไหม?
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
AgentWire ใช้ทรานสปอร์ตแบบใด?
AgentWire ทำงานแบบโลคัลผ่าน stdio
AgentWire ถูกตรวจสอบล่าสุดเมื่อใด?
ตรวจสอบล่าสุดเมื่อ 2026-10-28 OpenTelemetry + Wireshark + Cloudflare for AI Agents. Monitor, inspect, secure, replay, and optimize all traffic between Users, AI Agents, LLMs, and MCP Servers.