cwi-mcp-server
OfficialThis is a read-only MCP server that provides seven tools for inspecting CWI's trust infrastructure and verifying data integrity, without any write or mutation capabilities.
Ledger State & Version: Get the current Gear Ledger version (
ledger_state_version), full state or selected top-level fields (ledger_state_get), and list all agents with presence heartbeats (ledger_agents).Task Management (read-only): List task summaries with optional state filtering (
ledger_tasks) and fetch full task details by ID, including state history and artifacts (ledger_task_get).Trust Scoring: Use the
trust_verdicttool to score agent trust deterministically based on provided evidence; it returnsinsufficient-datawhen evidence is weak, never inventing a score.Needle Drop Verification: Verify the hash-chain integrity of a NEEDLE DROP ledger (
needledrop_verify), either the default example or any specified ledger file, detecting tampering.Read-Only Guarantee: All tools are strictly read-only, with no signing, no state mutation, and no secrets. The server holds no credentials.
Local & Public Data: Ledger reads work from CWI's public gear-ledger repo (no auth) when the local CLI is absent; trust and needledrop verification are fully local and offline.
Protocol Support: Works over stdio or as a streamable-HTTP bridge (for clients like Muse), with rate limits and body caps enforced over HTTP.
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., "@cwi-mcp-server Check the trust verdict for agent agent-42 with the latest signals."
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.
cwi-mcp-server
CWI's read-only MCP (Model Context Protocol) server. Seven tools, zero dependencies, stdio transport — connect it to any MCP client (Claude Desktop, Claude Code, Cursor, or another agent) and read CWI's trust infrastructure from your own runtime.
What you get:
Gear Ledger reads (5 tools) — the live, public provenance log of the CWI agent company: version, full state, agents + presence, task summaries, single-task detail.
trust_verdict— score agent trust with the CWI Verdict Engine v1.0.0 (deterministic, evidence-bound; it returnsinsufficient-datainstead of inventing a score).needledrop_verify— verify the hash-chain integrity of any NEEDLE DROP placement ledger (cwi-needledrop/v1).
Read-only means read-only. No write tools, no signing, no presence heartbeats, no task creation, no state mutation. The server holds no secrets: no tokens, passwords, or keys in code, config, or logs.
Don't trust us — see VERIFY.md for how to check every claim yourself, cold, in under five minutes.
Install (copy-paste)
Requirements: Node ≥ 18 and python3 on your PATH. Nothing to install — there are zero dependencies.
git clone https://github.com/CumulativeWebInc/cwi-mcp-server.git
cd cwi-mcp-server
node test.js # expect: 28/28 tests passed
node test-http.js # expect: 13/13 HTTP tests passedThat's it. server.js is the server.
New here? Start with the 10-step equip guide (every step tested): https://cumulativewebinc.github.io/cwi-learn/EQUIP-GUIDE.md
Related MCP server: cluby-mcp
One-click install
Install in Cursor — opens Cursor with the
cwiMCP config pre-filled (stdio:node /absolute/path/to/cwi-mcp-server/server.js). Replace the path with your real checkout path.Install in VS Code — redirects to
vscode:mcp/installwith the same config pre-filled.
Connect your MCP client
Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"cwi": {
"command": "node",
"args": ["/absolute/path/to/cwi-mcp-server/server.js"]
}
}
}Claude Code / any stdio MCP client: same shape — command node, one
argument: the absolute path to server.js. Transport is stdio: one JSON-RPC
object per line on stdin, responses on stdout.
HTTP bridge (for Meta Muse custom connectors and other HTTP MCP clients)
server-http.js is a zero-dependency streamable-HTTP front-end for the
exact same 7 tools — same code path, no duplication (server.js's
handleMessage() answers every request). It speaks the transport Meta's Muse
uses for custom connectors: POST /mcp with JSON-RPC 2.0, stateless (no
session id required), 202 on notifications, 405 on GET /mcp.
node test-http.js # expect: 13/13 HTTP tests passed
PORT=3000 BIND=0.0.0.0 node server-http.js
# -> cwi-mcp-server-http v0.3.0 listening on 0.0.0.0:3000 (POST /mcp)Host it on any always-on machine with a public HTTPS URL (Railway, Render,
Fly.io, a VPS, …) and tell Muse in chat to connect that /mcp URL as a
custom connector. Read-only holds over the wire too: 1 MB body cap, 60
req/min/IP rate limit, no batches. GET / serves a human info page,
GET /health a JSON health check.
Hosting (Docker-ready)
Dockerfile (Node 20 + python3) and render.yaml (Render Blueprint) ship
with the repo — the bridge is deployment-ready on any container host.
Render (free, no card): in Render, New → Blueprint → connect this repo. The blueprint deploys the Docker service with
/healthchecks. Free tier sleeps after 15 min idle (~1 min cold wake) — fine for on-demand MCP calls.Any Docker host (VPS, Zeabur, Cloud Run, …):
docker build -t cwi-mcp .then run with-p 7860:7860; the MCP endpoint ishttps://<host>/mcp.Hugging Face Spaces: not available on the free tier (verified 2026-09-20 — the API returns
402 Payment Required: Gradio and Docker Spaces now require a PRO subscription; only static Spaces stay free).
The 7 tools
# | Tool | Arguments | Returns |
1 |
| none |
|
2 |
|
| Full ledger state, or selected top-level keys |
3 |
| none | Every registered agent + latest presence heartbeat |
4 |
|
| Task summaries; filter by lifecycle state |
5 |
|
| Full task detail incl. state history and artifacts |
6 |
|
| Trust score or honest |
7 |
|
|
|
Example — read the ledger version
// tools/call {"name": "ledger_state_version", "arguments": {}}
{
"version": 1541,
"updated_at": "2026-09-17T11:23:35Z",
"sha": "3b2b46c8b8e14e0a3351e8896c6bd75e53391cde",
"tasks": 38,
"agents": 11
}Example — score trust (or get an honest refusal)
// tools/call {"name": "trust_verdict", "arguments": {"input": {
"engine_version": "1.0.0",
"subject": {"agent_id": "some_agent"},
"context": "agent-trust",
"observed_at": "2026-09-17T12:00:00Z",
"signals": {"erc8004": [], "needle_drop": [], "first_spin": []}
}}}
{
"status": "insufficient-data",
"score": null,
"missing": ["at least 3 verified signals across 2 families"],
"input_sha256": "9f2c…"
}Empty evidence → insufficient-data, never a made-up number. That's the
engine's whole point. Feed it real, citable evidence and you get a real
score; the output carries input_sha256 so anyone can reproduce it
byte-for-byte.
Example — verify a NEEDLE DROP ledger
// tools/call {"name": "needledrop_verify", "arguments": {}}
{
"file": "vendor/needledrop/example-ledger.json",
"ok": true,
"messages": ["chain intact"]
}Point file at any absolute path to a cwi-needledrop/v1 ledger to verify
that one instead. Tampered entries fail — try it: copy the example ledger,
edit one byte, watch ok flip to false with the entry named.
How the ledger reads work on your machine
On CWI's infrastructure the tools read through the canonical ledger CLI. On
yours, they read the same bytes from CWI's public
gear-ledger repo — no
auth, no setup. (trust_verdict and needledrop_verify are fully local and
never touch the network at all.)
Files
server.js— the server (7 tools, stdio, zero deps)server-http.js— streamable-HTTP bridge (statelessPOST /mcp, zero deps)test.js— full protocol + tool harness (node test.js→ 28/28)test-http.js— HTTP bridge harness (node test-http.js→ 13/13)VERIFY.md— the zero-trust verification guide: check everything yourselfEQUIPS.md— public, receipt-only log of external equipsagent-card.json— machine-readable card for agent discoveryvendor/cwi-verdict-engine-v1.0.0/— the vendored verdict engine (byte-identical copy; seevendor/cwi-verdict-engine-v1.0.0/SOURCE.md)vendor/needledrop/—ledger.py+ schema + a 2-entry example ledger (entries sealed by the realledger.py, clearly labeled as examples)examples/— real verdict output from a 2026-09-17 run
Result, measurement, kill rule
Result this must produce: an external agent calls a tool with their identity attached. Receipts go in EQUIPS.md.
Measured by: real tool calls with checkable receipts.
Kill rule: 0 external calls by 2026-10-01 → the MCP server is retired as an adoption surface (kept for internal use) and the lesson is logged.
License
MIT — see LICENSE.
Available Tools
7 toolsledger_agentsA
Read-only: Gear Ledger agents with their latest presence heartbeat merged in (status, last seen, current task, note). No arguments. On machines without the CWI ledger CLI, reads the public CWI gear-ledger repo (no auth, read-only).
Inputs: no named parameters.
ledger_state_getA
Read-only: full Gear Ledger state.json document (agents, tasks, presence, handoffs, approvals). Optionally pass fields to select top-level keys. On machines without the CWI ledger CLI, reads the public CWI gear-ledger repo (no auth, read-only).
Inputs: fields.
ledger_state_versionA
Read-only: Gear Ledger version summary (version number, updated_at, repo sha, task count, agent count). No arguments. On machines without the CWI ledger CLI, reads the public CWI gear-ledger repo (no auth, read-only).
Inputs: no named parameters.
ledger_task_getA
Read-only: full detail of one Gear Ledger task by task_id (including state history and artifacts). On machines without the CWI ledger CLI, reads the public CWI gear-ledger repo (no auth, read-only).
Inputs: task_id (required).
ledger_tasksA
Read-only: list Gear Ledger tasks (summaries). Optionally filter by state (created, assigned, in_progress, delivered, verified, cancelled, failed). On machines without the CWI ledger CLI, reads the public CWI gear-ledger repo (no auth, read-only).
Inputs: state.
needledrop_verifyA
Verify-only: check the hash-chain integrity of a NEEDLE DROP placement ledger (cwi-needledrop/v1). Recomputes every entry hash, checks prev_hash linkage, chain head, and revision-note consistency. Tampered or corrupt ledgers fail. No signing, no sealing — verification only. Defaults to the vendored example ledger (vendor/needledrop/example-ledger.json); pass an absolute file path to verify any other cwi-needledrop/v1 ledger, e.g. CWI's canonical needle-drop.json.
Inputs: file.
trust_verdictA
Score agent trust with the CWI Verdict Engine v1.0.0 (deterministic, evidence-bound). Pass input as the engine input object: {engine_version, subject:{agent_id,...}, context, observed_at, signals:{erc8004:[], needle_drop:[], first_spin:[]}}. The engine NEVER invents a score: insufficient evidence yields status "insufficient-data" (score null), disputed evidence yields "evidence-disputed" — report those as-is. Inputs must be real, citable evidence.
Inputs: input (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 7 tools
Each tool has a distinct purpose: ledger_state_get returns the full state document, ledger_agents and ledger_tasks are specialized views, ledger_task_get is a detail lookup, and trust_verdict/needledrop_verify cover different verification domains. The only mild ambiguity is that ledger_state_get can also reach agents/tasks via its fields argument, but descriptions make the specialized tools clearly preferable for those cases.
The ledger tools share a useful `ledger_` prefix and snake_case, but verb placement is inconsistent: some names end with `_get` (ledger_state_get, ledger_task_get), others are bare nouns (ledger_agents, ledger_tasks), one is a noun phrase (ledger_state_version), and the non-ledger tools use different forms (trust_verdict, needledrop_verify). The convention is readable but not predictable enough to infer tool names without reading descriptions.
Seven tools is a well-scoped set for this server. It provides the core read-only ledger views, a task detail endpoint, the state version, plus two specialized verification/scoring tools without unnecessary redundancy.
The read-only ledger surface covers the main entities: full state, agents with presence, task summaries, and task detail, plus a version endpoint. The main gap is that handoffs and approvals are only reachable through ledger_state_get rather than dedicated tools, but agents can work around that by requesting the full state or selecting fields.
Related MCP Connectors
Verifies Bernstein run receipts and hash chains; lists the shipped presets and adapters. Read-only.
Read-only gateway for durable agent identity, consent, recognized work, and signed receipts.
Read-only verifier for 25 ProofRelay MCP tools and non-confidential evidence bundles.
Issue & verify signed (ed25519), hash-chained, timestamped provenance receipts for agent actions.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables querying enterprise records and retention policies from any MCP client over stdio, with read-only tools for searching records, fetching retention verdicts, identifying archival candidates, summarizing departments, forecasting retentions, and viewing audit history.-
- AlicenseAqualityCmaintenanceEnables agents to read Cluby markets, positions, and pre-trade quotes over stdio, using the same arithmetic as the interface while holding no keys and signing nothing.9MIT
- AlicenseNot gradedqualityAmaintenanceEnables agents and applications to record, query, verify, and receive hash-chained event provenance over stdio, supporting self-metering and independent audit of autonomous system activity.6MIT
- AlicenseAqualityBmaintenanceExposes forensic case and evidence management operations over stdio, including case creation, evidence registration with SHA-256 hashing, and integrity verification, so clients can manage DFIR investigations programmatically.7MIT