Implements OpenAI MCP Events as the persistent signed event protocol, enabling durable event delivery and receipts for compatible wake-up runtimes.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Kadenco inspect native T3 inventory and list recent signed receipts"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Experimental standalone release candidate, 0.1.0-alpha.1. No T3 source edits.
MCP tools for native T3 inventory/results, persistent signed events, explicit bounded delegation and durable receipts. Not a replacement agent, orchestrator, queue or memory system. Existing Eddie tool/environment names are retained as compatibility identifiers; they do not require an OpenAI dot. A persistent agent needs an MCP client plus a compatible wake-up runtime. OpenAI MCP Events is the implemented event protocol; other runtimes need compatible webhook ingress or explicit read-only polling. Basic MCP tool support alone is insufficient.
Use Node 22.18+ or 24.11+. Run npm ci, then npm exec -- vp run test. Vite+ 1.4.0 runs the existing node:test suite; vp test is the separate Vitest command and is not used here. Start with npm exec -- vp run start. Copy examples/read-only.json to config.local.json, set your own absolute T3 database path, and create private-state with owner-only permissions (chmod 700 on POSIX; an owner-only ACL on Windows). Do not use the example zero hash for a real principal: generate your own random bearer and store only its SHA-256 in policy. Supply EDDIE_T3_CONFIG as the absolute policy path and EDDIE_T3_STATE_KEY as your own securely provisioned base64 32-byte key, then npm exec -- vp run start. Keep key and bearer out of shell history/logs.
The listener binds loopback on port 41880. Connect an MCP client to /mcp using its owner-provisioned bearer. The minimal example has no enrolled sessions, event destinations, execution grants or guest access; configure exact sessions and read scopes for your own T3 instance before use. Nothing is auto-enrolled or discovered across trust boundaries.
Public startup uses read-only local durable projections and optional configured read routes. It does not load a private fleet helper. Legacy fixed verify_result/helper-backed startup profiles are not supported by this public entrypoint. Native commands are separate opt-in policy, not enabled by the example. The broker uses explicitly configured native CLI authorization; never treat subscriptions as execution authority. Secure HTTPS/private connectivity and external identity enrollment are operator responsibilities; no MTG/Keeper deployment is required or bundled.
Related MCP server: attest-mcp-server
Read COMPATIBILITY.md and SECURITY.md. Native APIs/schema can change. Strict active-run steering is disabled: installed native behavior can start/queue a turn in races. continue_thread explicitly authorizes a new turn. Stored, native accepted, observed, agent-acknowledged and completed are distinct. Reconcile a stable UUID after uncertainty; never automatically resend. HTTP 200 webhook delivery does not prove agent execution. Aggregate monitoring currently requires live command scope as well as read scope. Diagnostics retain seven days or 25,000 records globally. Native completion observation retains 20 recent runs; this is not lossless telemetry.
Collaboration remains opt-in and disabled by default. The current identity verifier requires a configured Entra/Access identity boundary. Do not assume arbitrary external guest admission. Sharing links grant no authority.
Experimental GitHub source release. The private package flag deliberately prevents npm publication; install from source. Project code is AGPL-3.0-or-later, selected by the owner. Project: https://github.com/Midtown-Technology-Group/kadenco . No real fleet credentials, runtime files, private Git history or deployment evidence are included. The export manifest binds every candidate file; install dependencies separately from the lockfile.
This server cannot be deployed
Agent-native MCP for governed commerce, x402 payments, paid capabilities, and verifiable receipts.
Public MCP for agent verification, work discovery and governed interoperability.
Read-only Remote MCP for externally grounded AI agent trust receipts.
Governed MCP: agent audit, provenance, deterministic checks, and receipt-backed FragGate execution.