Continuity Evidence

io.github.zerohourzuluPublished by GitHub account zerohourzuluv0.3.0-preview.12Updated Oct 24, 2026

A bounded local evidence tool with current agent permission checks and durable retry handling.

VerifiedSTDIODesktop onlyDeveloper ToolsSecurity & Monitoring

Overview

AI-generated overview

A local MCP evidence tool that records agent roles, permissions and unfinished duties across agent replacements, with durable retry handling.

What you can do
Record which role an agent occupies, which powers it holds and what happened before a replacement, so unfinished duties survive the change while permissions are re-checked independently. Produce and inspect evidence for a case, including refusals that create no packet, and explain saved records. Keep case history locally or, in a chain-backed deployment, make selected history publicly verifiable through cryptographic commitments. Embed the engine through a supplied JavaScript SDK with TypeScript declarations and an offline consumer check. Connect cooperating remote tools and the MCP evidence tool to one shared Core dependency.
When to use it
Useful when an assistant or agent workflow must hand over between agents and you need a local, inspectable record of who holds a duty and what the successor may do. Also suited to evaluating or integrating permission checks and evidence recording into a framework or policy engine.
Requirements
Runs as a local stdio process from the npm package. The tutorial requires Node 22.18+ (22.x), 24.x or 26.x on macOS or Linux; no global pnpm or Python is needed. Setup downloads locked dependencies with scripts disabled. No account is needed for the public walkthrough; the manifest declares no authentication, environment variables or headers.
Before you install
The README states this is local reference software, not a hostile-agent sandbox, production credential system or autonomous cyber defender, and that real integrations must mediate every consequential operation through a protected executor. Chain-backed deployments carry their own privacy, availability and finality requirements, and cryptographic commitments can let authorized reviewers check private records. Setup downloads dependencies, and rerunning the tutorial creates new cases while keeping

Installation

In SourceWeft

  1. Open Continuity 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

Message package preview — 10 October 2026: Core11, remote6, gateway11 and evidence12 add the optional saved-record explanation command. See the current release guide. Existing API/policy/proof behavior stays unchanged; older npm archives remain available.

Source preview.13: MCP client/server/core updated to 2.2.0 for GHSA-6qxp-vccf-f47h. New MCP archives are source-prepared; npm and Registry still select the earlier published tuple.

Previous source preview.12.1: Plain-language messages, completed historical ZK experiments, and retained-receipt reader source. The installable packages now include the optional saved-output explanation command; use the exact tuple in the release guide.

Current developer preview: Exact versions and installation.

Follow an investigation: Follow one protected case through interruption, replacement, a signed completion, new evidence and a persistent challenge. Run the walkthrough.

Continue an existing case: Same case, longer history, preserved permissions and unfinished duties. Start with supported profiles, migration and recovery, or the protected evidence example.

Core 0.3 preview: Build a fresh case, connect approved MCP tools, or run the public tests. The original 0.2.2 evaluation releases remain available.

[it is 2 a.m. do you know what your agent is doing?] [Security. Control. Inheritance. Durable survival of powers and responsibilities. Record durably, locally or on chain. This is your agent’s brain on Continuity.]

Continuity — agents change; responsibility remains

Explore the interactive demo → · Start here · Developer guide · Security and limits

Try your own choices in the playground → No installation or account needed.

Developer preview · Apache 2.0. Release status.

A process supervisor can start Agent B. Continuity records which role B occupies, which powers it has, what happened before, and which unfinished duties survive the change. A replacement receives its explicitly granted powers; an unfinished duty does not silently grant more authority.

The original tutorial follows a security investigation interrupted by an agent replacement. The old request is refused, the investigation stays OPEN, and the successor’s review permission is checked independently.

Choose your path

You want to…Start hereWhat you will learn
Understand the idea without installing anythingThe story and answer keyWhy an unfinished duty and permission are separate.
Run and inspect a replacementTutorial, then operator commandsWhat was refused, who holds the duty and what B may do.
Evaluate or integrate the codeCore 0.3 quickstartCreate your own permissions, try a protected tool, then connect a framework or policy engine.

Run the tutorial

Use Node 22.18+ (22.x), 24.x or 26.x on macOS or Linux. Prefer the latest patch of your chosen release. No global pnpm or Python is needed for this tutorial. Full quickstart, including archive extraction.

Get the public source; no GitHub account is needed:

sh
git clone https://github.com/zerohourzulu/continuity.gitcd continuity

See the result first, without installing dependencies:

sh
node examples/recorded.mjs

Expect ALLOW for reviewing the investigation and DENY for collecting another packet. This inspects a bundled recorded case; it does not execute a new action.

Generate your own evidence:

sh
node tools/setup.mjsnode tutorial/start.mjs

Expect PASS — duty remains OPEN; B has no collection power, then VERIFIED. Setup downloads locked dependencies with scripts disabled. The tutorial then runs locally using synthetic data and public test keys. It prints your new case name and exact follow-up commands. Rerunning creates another case and preserves earlier evidence.

Change one permission:

sh
node tutorial/start.mjs --deny-review

B still inherits the duty, but its review permission becomes DENY. Expected output and explanation · Setup checks and recovery.

Choose your holder of record.

Keep records locally, or choose a chain-backed deployment to make selected history publicly verifiable. Cryptographic commitments can let authorized reviewers check private records without publishing their contents. This tutorial runs locally; each deployment has its own privacy, availability and finality requirements.

Explore and build

Your next stepWhere to go
See the idea and recorded comparisonInteractive demo
Inspect your new run and optional recorded Linux evidenceOperator guide
Integrate an authority checkDeveloper guide · Executable reader example
Try optional native enforcement mechanicsDisposable Linux lab
Review source provenance and validationValidation · Presentation checks · Provenance
Extend the presentationWebsite maintenance

Presentation maintainers can find the separate owner-private edition in website maintenance. The public walkthrough needs no account.

Contribute and explore

Read the short development history from Core 0.1 to the current reference implementation. Contributions and public RFCs are welcome; an RFC is useful for changes to protocol behavior or security boundaries.

Scope

Core 0.2.2 is local reference software. It is not a hostile-agent sandbox, production credential system or autonomous cyber defender. The website selects recorded results; the downloaded tutorial executes Core. Real integrations must mediate every consequential operation through a protected executor. Security boundaries.

The source, tutorial, tests, reader and website are included here and in the complete evaluation download. Earlier editions remain available as historical releases; use evaluation.6 for the commands above. Package overview · Presentation provenance.

Release status and licensing · Apache 2.0 license · Third-party notices

Embed the engine

Try the standalone developer package: install the supplied JavaScript SDK tarball in a separate application, create one authorized synthetic packet, and inspect why refused requests create none. Includes TypeScript declarations and an offline consumer check. This is an evaluation interface around unchanged Core0.2.2.

Shared Core integration preview

The shared-package preview connects the cooperating remote tools and MCP evidence tool to one Core dependency. Try native LangChain tools and lost-reply recovery, or the MCP evidence tool.

Source: README.md at commit 312b3f1

Tools

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

Version history

4
  1. v0.3.0-preview.12LatestOct 11, 2026
  2. v0.3.0-preview.11Sep 26, 2026
  3. v0.3.0-preview.9Sep 23, 2026
  4. v0.3.0-preview.8Sep 23, 2026

Score

50% 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