Repository avatar
Security
v0.3.3
active

semantic-state-runtime

io.github.oaslananka/semantic-state-runtime

Local MCP server for policy-authorized semantic context and source-authority reconciliation.

Documentation

Semantic State Runtime

CI npm release

GPL-3.0-only license GitGem

GitHub Sponsors Buy Me a Coffee Stars

Switch AI coding agents without losing the current project facts.

Connect one public GitHub repository or selected local folder, then ask your MCP-capable AI agent for the open issues and relevant decisions. SSRL keeps a bounded, read-only project view updated on demand and applies explicit source authority instead of trusting stale local notes.

Status: public npm MCP release (current version in the npm badge). The MCP runtime is usable with synthetic data, but independent human onboarding and production use are not yet validated. No hosted service or model account is included.

Semantic State Runtime (SSRL) helps AI agents use the right information when several sources disagree. It reconciles writable state against explicit source authority, persists the result across sessions and exposes only policy-authorized context through the Model Context Protocol (MCP).

Why it matters

Imagine a project deadline is November 15 in a stale local copy, but the trusted source says November 20. Rather than letting an AI assistant guess, SSRL can:

  1. Identify which source is authoritative for the field.
  2. Produce and apply a precondition-checked correction to the writable copy.
  3. Give a new MCP client the authorized, current deadline after restart.
  4. Exclude fields and evidence the client is not permitted to see.

This is state reconciliation and access-controlled context, not simply chat-history memory, an unrestricted file search, or an LLM provider.

Connect your first real project

npm install -g semantic-state-runtime
ssrl-connect https://github.com/oaslananka/semantic-state-runtime
cd ssrl-semantic-state-runtime
opencode

Ask: "Use SSRL context.compile for oaslananka/semantic-state-runtime. Which GitHub issues are currently open? Show issue numbers and sources." The generated OpenCode configuration is ready; Antigravity's workspace-local .agents/mcp_config.json and a separate Copilot MCP entry are included.

Alternatively, connect a local folder once using ssrl-connect ./my-project --out ./my-project-ssrl. It discovers the project's GitHub origin when present, or accepts --local-only. Existing files are never modified; only selected Markdown, and GitHub open issue/PR metadata when connected, are ingested. The bounded refresh runs when context is requested, not through a background service. This is not an automatic history of every chat or file.

See the one-command project connection guide for exact limits, private GitHub access, refresh timing, source authority and reproducible acceptance evidence.

Install from npm

Requirements: Node.js 24.12+ and npm. No Git clone, workspace build, API key, or model account is needed to try the synthetic read-only MCP server.

npm install -g semantic-state-runtime
ssrl-demo-init ./ssrl-example
ssrl-node --version

The initializer creates a new folder and refuses to overwrite one. It writes read-only synthetic Markdown and ready-to-use client definitions:

  • OpenCode: open ./ssrl-example as the working directory; the generated opencode.json uses the installed server's absolute paths.
  • VS Code Copilot: in Copilot, choose Chat: Open Customizations → MCP Servers → Copilot Global. Copy the ssrl entry from ./ssrl-example/copilot.mcp-config.json into your existing global mcp-config.json (merge; do not replace other configured servers).

The underlying ssrl.config.json limits access to synthetic records and does not grant writes. Ask the client to call context.compile for Project Atlas deadline; the authorized response contains 2026-11-20. A model account may be needed by your chat client, not by SSRL itself. Client definitions are generated inside the demo folder; your global configuration is never modified.

Use the explicit configuration guide before authorizing access to real files. Windows and Linux registry installs plus MCP tool calls passed with synthetic data; independent human onboarding and production suitability remain unverified.

See installation and client setup.

Run the demo

Requirements: Node.js 24.12+ and pnpm 12.14.0. From a fresh clone:

pnpm install --frozen-lockfile
pnpm build
pnpm demo:handoff

The demo uses synthetic Markdown/SQLite data and two actual MCP SDK client sessions. Its output shows the stale value, trusted value, state.plan/state.apply reconciliation, restarted runtime, authorized context.compile answer, and privacy checks. Package installation may use the network; the demo itself requires no API key or external model.

Walk through the real observed output.

Use it as an MCP server

For a source-built alternative, see the source-based Developer Preview installer. It builds a production-only installation at an explicit final path; the source-based Windows installation cannot safely be moved afterward. The published npm package is standalone, separate from that limitation.

Maintainers can still build a private local npm tarball using pnpm preview:pack -- <new-staging-directory> and verify it with pnpm test:preview-pack. The stable npm package is publicly available with GitHub release provenance. The Registry listing must be verified separately against the official MCP Registry API; independent consumer acceptance has not yet been established.

When running a local source build instead, after creating an explicit configuration file using the local runtime guide, start the server with:

node packages/node-app/dist/cli.js ./ssrl.config.json

./ssrl.config.json is your own file, not an included ready-made production config. Use explicit source mappings, principal scopes, field policies and local paths. The server does not silently discover user files. Point your MCP-capable client's stdio command and arguments at the built CLI and that config; each chat client has a different config format.

MCP toolWhat it does
state.planComputes a bounded proposed change based on source authority and preconditions.
state.applyApplies an approved proposal only within configured write permissions.
context.compileReturns current, policy-scoped semantic context within a bounded token budget.

Start read-only. Grant only state:read and context:read when a client needs retrieval, and disable writable field policies and plan/apply. The first-run OpenCode setup creates such a temporary synthetic read-only workspace automatically:

pnpm demo:chat-setup

The setup is offline; an actual hosted-model chat request requires a separate, explicit developer API key and network access.

Verified AI client integrations

OpenCode v2.3.0 and Goose CLI v1.59.0 each made a real context.compile MCP call and returned the correct authorized deadline from synthetic data. They are two independently tested chat applications, not two independently hosted model providers: both used the same NVIDIA developer gateway on one Windows host. These tests do not establish physical cross-device synchronization or readiness for real user data.

See OpenCode and Goose examples and the separate live two-model handoff.

Project status

  • Ready to try: public npm stable release (latest), read-only synthetic first-run, local build, deterministic demo, stdio MCP integration and regression tests.
  • Not yet delivered: independently validated human onboarding or a fully supported consumer product. Registry discoverability is verified separately through the official public API after publication.
  • License: GNU GPL v3.0, copyright (c) 2026 Osman Aslan. Redistribution is subject to the GPL copyleft terms.
  • Participation: read Contributing and Security. Security issues can be reported privately.

Architecture and implementation

source systems / local files
        ↓
connectors + restart-safe ingestion
        ↓
immutable Artifact Plane + Semantic State Store
        ↓
bitemporal state / retractions / typed identity / relations
        ↓
incremental Context Capsules
        ↓
Context Access Gateway
        ↓
retrieval / future agent adapters

AI side effects:
agent -> RuntimeHost plan/apply -> provider preconditions -> journal

Implemented checkpoints include:

  • canonical entity identity and typed alias resolution;
  • bitemporal semantic observations and immutable retractions;
  • explicit field authority and conflict detection;
  • deterministic reconciliation planning and proposal digests;
  • durable SQLite semantic state + ordered change feed;
  • incremental, time-aware context capsules;
  • Markdown filesystem and Google Calendar ingestion adapters;
  • immutable content-addressed Artifact Plane;
  • policy-enforced artifact access;
  • local stdio MCP composition;
  • authenticated stateless MCP 2026-07-28 HTTP composition;
  • principal-aware, capsule-backed derived context access;
  • MCP context.compile projection with authenticated context:read preflight;
  • policy-scoped prefix-Merkle replication with bounded resumable sessions;
  • principal-scoped immutable record + artifact blob transfer;
  • authenticated bounded replication HTTP with real two-node semantic/artifact convergence;
  • optional Ed25519 device-bound HTTP Message Signature authentication with replay defense;
  • epoch-scoped opaque encrypted replication object identities that hide plaintext record/blob identifiers and digests from an untrusted relay;
  • descriptor-generic prefix-Merkle + bounded reconciliation with epoch-scoped opaque views and no plaintext reconciliation metadata.

Design posture

Core constraints are intentionally stricter than a conventional "RAG over personal files" system:

  • source evidence remains auditable;
  • derived caches/indexes are rebuildable and never define truth;
  • authorization happens before identity resolution/ranking when disclosure could leak hidden state;
  • agents receive bounded semantic context rather than unrestricted source-system search;
  • transport authentication and domain authorization are separate layers;
  • protocol adapters such as MCP do not define canonical semantics.

See docs/invariants.md and the versioned design documents under docs/.

Near-term gaps

The runtime has a public stable release and a one-command read-only project connector, but independent novice-user onboarding and private-data production readiness remain unverified. Important remaining work includes:

  • more source connectors and semantic mappings;
  • historical context-access semantics;
  • opaque provenance/evidence expansion handles;
  • device enrollment, sender-constrained authentication and key rotation;
  • ciphertext object storage and opaque-only HTTP relay integration for untrusted relay/cloud deployments;
  • active-active source receipt/witness semantics;
  • deployment composition, real IdP verifier integration and observability;
  • product UX, onboarding, sync operations and commercial packaging.

Commands

pnpm install --frozen-lockfile
pnpm typecheck
pnpm test
pnpm build
pnpm benchmark:context:v2
pnpm benchmark:google-calendar:v1
pnpm benchmark:markdown:v1
pnpm test:stdio