TollWire Evidence

io.github.TollWirePublished by GitHub account TollWirev0.1.2Updated Nov 7, 2026

Tamper-evident evidence log of agent changes, scans and approvals, mapped to ISO 27001/NIS2.

The publisher's own server: Wire Cash: bank wire fees and transfer costs

VerifiedSTDIODesktop onlyDeveloper ToolsSecurity & Monitoring

Overview

AI-generated overview

Local MCP server that records agent changes, scans, approvals and deployments in a tamper-evident hash-chain log and verifies it.

What you can do
  • Append events such as changes, reviews, scans, approvals, deployments, incidents and permission changes to a hash-chain log, with a filter that refuses recognized credential formats
  • Read and filter logged events, and verify the hash chain plus any local or external anchors passed in
  • Create an anchor of the latest entry and return a suggested signed Git tag command
  • Map event types to ISO/IEC 27001:2022 Annex A and NIS2 Article 21(2) controls (draft)
  • Export a Markdown evidence report with integrity and coverage
When to use it
Useful when you want a local, append-only record of what an AI agent changed, approved or deployed in a project, and a way to check later that the record was not silently rewritten. Suited to teams preparing audit evidence, provided they understand that only externally stored anchors give real protection.
Requirements
Runs locally over stdio via npx from the npm package; Node.js 22.12.0 or later. No authentication, API keys or network access. The log path comes from TOLLGATE_EVIDENCE_LOG or defaults to .tollgate/evidence/log.jsonl under the project directory; the anchor file sits next to it. Paths must not contain symbolic links, and files must be regular files.
Before you install
The hash chain detects changes but does not prevent them; anyone who rewrites the whole file can build a new valid chain, and entries after the last external anchor are unprotected. Actor names and anchor events are self-declared and unverified, and the server never signs anything. The control mapping is a draft not confirmed by an auditor. The secret filter only catches fixed-prefix credential formats, so do not put secrets in events. Version 0.x interfaces may change incompatibly.

Installation

In SourceWeft

  1. Open TollWire Evidence in the dashboard and add it to a workspace.
  2. 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

TollWire Evidence MCP (tollwire-evidence, MVP)

MCP server that records agent changes, reviews, scans, approvals, deployments, incidents and permission changes as a tamper-evident hash chain. It maps event types to ISO/IEC 27001:2022 Annex A and NIS2 Article 21(2) controls (draft).

Status: version 0.x (MVP). Interfaces can change incompatibly before 1.0. Installation, the marketplace and npm options, and the project-wide security notes are in the top-level README. Issues are welcome; pull requests from third parties are not accepted, see CONTRIBUTING.md.

Tools

ToolPurpose
record_eventAppend an event. Events that contain a recognized secret format are refused. Type anchor is accepted only with a matching anchor (see below).
list_eventsRead events, optionally filtered by type.
verify_logCheck the hash chain. Also checks the local anchor file (if present) and any external anchors passed in anchors.
create_anchorWrite the last entry as an anchor {seq, hash, ts, count} to the anchor file. Returns the anchor line, a tag name and a suggested signed Git tag command. Writes no event.
map_controlsMap event types to controls (draft, not confirmed by an auditor).
export_reportMarkdown evidence report with integrity and coverage.

Storage

  • Log: TOLLGATE_EVIDENCE_LOG if set; otherwise .tollgate/evidence/log.jsonl in the project directory (CLAUDE_PROJECT_DIR, or the current directory).
  • Anchor file: <log>.anchors.jsonl in the same directory. Append-only. Each line carries prev_anchor_hash and anchor_hash, so the anchor file is itself tamper-evident.
  • Lock: a lock file next to the log.
  • Permissions: files are created with mode 0600.
  • Path: every path component between the root and the log must be a real directory, not a symbolic link. Otherwise all tools that access files refuse to work. The root itself may be reached through a link.
  • Files: log and anchor file must be regular files with exactly one directory entry. Symbolic links and hard links are refused. Checks are repeated on the opened file descriptor.
  • Errors: tool error responses contain no stack trace and no file paths. Unexpected errors are reported generically.

External anchors

An anchor records the sequence number, hash, timestamp and count of the last entry in canonical JSON, for example {"count":42,"hash":"…","seq":42,"ts":"2026-10-04T08:00:00.000Z"}.

verify_log checks, besides the chain, that for each anchor an entry with that seq exists and carries exactly that hash and timestamp. A log that was rewritten after the anchor (a new, internally valid chain) or truncated is then detected.

  • valid means only "internally consistent".
  • anchored is true only if the result is valid and at least one external anchor was confirmed. valid: true with anchored: false is not proof that the log is unchanged.
  • A local anchor file does not protect against someone who rewrites or deletes the log and the anchor file together. An anchor is effective only when it is stored outside the machine.

Recommended procedure, performed by a person with their own signing key:

  1. create_anchor (or node scripts/anchor.mjs create in the package directory).
  2. Store the anchor outside the machine, for example as a signed Git tag. The tools never sign anything.
  3. Record an anchor event with node scripts/anchor.mjs record '<anchor-json>' signed_tag <tag-name> <actor>. Exit code 0: recorded. Exit code 2: refused.

Verification: git tag -v <tag> checks the signature. Then pass the anchor line from the tag message to verify_log (anchors), or run node scripts/anchor.mjs verify '<anchor-json>' (exit 0 valid, 1 finding, 2 usage error).

anchor events

An anchor event documents that a person stored an anchor outside the machine. The control mapping is a draft (ISO/IEC 27001:2022 A.8.15, A.5.33, A.5.28; NIS2 Article 21(2)(f)).

  • Order: create_anchor, then the external storage, then the anchor event. Recording the event before storage would document a storage that has not happened yet.
  • The event carries ref (tag name or comment URL), details.anchor (the stored anchor) and details.storage (signed_tag or digest_comment).
  • The anchor must match the log: the entry at seq must have the same hash and timestamp. This is checked under the lock. Otherwise nothing is written.
  • The tool checks that the anchor matches the log. It does not check that the external storage happened, who stored it, or that ref points to a real storage location. Only the signed tag or comment itself is evidence of that.

Self-declared fields

The hash chain shows that the log is internally consistent. It does not show who wrote an entry.

  • actor is free text supplied by the caller. The server does not check identity. Any process that may write the log can record any actor, including the name of a person.
  • An anchor event is also self-declared. Only the external storage itself is evidence.
  • For reviewers, anchors in signed tags count, not anchor events in the log.

Secret filter

record_event refuses events whose content matches a known credential format: private keys (PEM, OpenSSH, PGP), GitHub, GitLab and npm tokens, AWS access keys, Anthropic and OpenAI keys (sk-…), Stripe keys (sk_live_, sk_test_, rk_…), Google API keys (AIza…), signed JWTs, Slack tokens and Slack webhook URLs. Only formats with a fixed prefix are recognized, which keeps false positives low. Unknown formats and freely chosen passwords are not detected. Do not put secrets into events.

Package contents

The files field of package.json includes dist/, .claude-plugin/, .mcp.json, scripts/anchor.mjs, LICENSE and README.md. Source code, tests, coverage and the smoke test are not shipped. Run npm run build before packing, so that no stale files are included.

Limits

  • The hash chain detects changes; it does not prevent them. Whoever rewrites the whole file can create a new valid chain.
  • External anchors detect this only for entries up to the last externally stored anchor, and only when the anchors are passed to verify_log. Entries after the last anchor are not protected.
  • Signing each entry and automated external storage are not implemented.
  • The control mapping has not been confirmed by an auditor and is not evidence of conformity with ISO/IEC 27001 or NIS2.
  • The server makes no network requests.

Development

bash
npm ci --ignore-scriptsnpm run check   # typecheck, lint, tests with coverage, build, smoke test

Node.js 22.12.0 or later is required (engines). Runtime dependency: zod only, pinned to an exact version.

Dependencies and security: the MCP server is a small built-in implementation (src/vendor/mcpStdio.ts): JSON-RPC 2.0 over stdin/stdout, one message per line, protocol versions as in the official SDK 1.32.0 (latest 2025-11-25). It opens no network connection, rejects lines over 1 MiB (the session ends with exit code 1) and the key __proto__, and writes diagnostics to stderr only. The official @modelcontextprotocol/sdk is a development dependency only: scripts/sdkCompat.mjs (part of npm run smoke) connects its Client with StdioClientTransport to the built server and calls every tool. The SDK and its HTTP stack are not installed with the package.

License

Source-available under the PolyForm Noncommercial License 1.0.0 (SPDX: PolyForm-Noncommercial-1.0.0), see also NOTICE. Noncommercial use only; commercial use requires prior written permission. Details: LEGAL.md.

Source: evidence-mcp/README.md at commit 3364077

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.2LatestNov 7, 2026

Score

40% 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

    Declare a license in the repository or the registry entry.

Other Wire integrations