1. Conduid
  2. RAG
  3. TheStack-ai/waypath
MCP server · RAG

TheStack-ai/waypath

aware recall via FTS5 and RRF, and governed memory with explicit promote and review gates. Ships Codex and Claude Code host shims and a native MCP server.

34Low

Scored 5 months ago · breakdown

About TheStack-ai/waypath

TheStack-ai/waypath is an MCP server in the RAG category: aware recall via FTS5 and RRF, and governed memory with explicit promote and review gates. Ships Codex and Claude Code host shims and a native MCP server.

Install

From source
git clone https://github.com/TheStack-ai/waypath

This server has no ConduID identity, so agent calls to it are not receipted. Pin the version you install and review the source before granting it credentials.

Ask AI

✦

Ask AI about TheStack-ai/waypath

Powered by Claude · Grounded in docs

I know everything about TheStack-ai/waypath. Ask me about installation, configuration, usage, or troubleshooting.

Security checks

  • ✓README presentThe repository ships documentation.

5 checks have not been run yet.

README

[!TIP] New here? The Quick start gets you from npm install to your first persistent agent session in about 60 seconds.


What is Waypath?

Waypath is a local-first knowledge engine for coding agents and solo developers. It stores your project decisions, entity relationships, and session artifacts in a single SQLite file, then serves graph-aware, truth-first context to any agent host — Claude Code, Codex, or an MCP client — through a thin CLI.

Unlike cloud memory services, Waypath:

  • runs entirely on your machine,
  • owns a canonical truth schema instead of a vector blob,
  • treats every memory as first-class with explicit promotion + review gates,
  • ships a 77 kB npm package with no required runtime services.

Why Waypath?

Problem Waypath's answer
Agents forget across sessions Persistent SQLite truth kernel
RAG returns irrelevant chunks FTS5 + RRF hybrid ranking with graph expansion
Memory services hallucinate silently Explicit page → promote → review governance
Cloud lock-in, data exfiltration Everything is one local .db file you own
Tool per host (Claude, Codex, Cursor) Single facade, thin host shims, native MCP server

Install

[!IMPORTANT] Requires Node.js ≥ 22. Node 22.5+ unlocks the native node:sqlite driver; earlier 22.x versions auto-fall back to better-sqlite3.

npm install -g waypath

Verify:

waypath --help
waypath source-status --json

Quick start

1. Bootstrap a session (Codex example):

waypath codex --json \
  --project my-project \
  --objective "ship v2 of the retrieval pipeline" \
  --task  "refactor hybrid ranker" \
  --store-path ~/.waypath/my-project.db

2. Recall relevant context:

waypath recall --query "hybrid ranker decisions" --json

3. Capture a distilled insight and promote it through review:

waypath page    --subject "hybrid ranker v2 design"
waypath promote --subject "hybrid ranker v2 design"
waypath review-queue --json

4. Run as an MCP server (for Claude Code, Cursor, any MCP client):

waypath mcp-server --store-path ~/.waypath/my-project.db

See it in action

$ waypath codex --json --project auth-service \
    --objective "migrate to passkeys" --task "design flow"
{
  "host": "codex",
  "session_id": "auth-service:passkey-flow",
  "context_pack": {
    "truth_highlights": {
      "decisions": [
        "Use WebAuthn level 2 with user verification required",
        "Argon2id for password fallback hashing"
      ],
      "entities": ["UserSession", "AuthGateway", "RefreshToken"],
      "contradictions": []
    },
    "recent_pages": [
      "Session storage design — promoted 2026-04-12"
    ]
  }
}

Command surface

Area Commands
Session bootstrap codex, claude-code, mcp-server
Recall recall, explain, graph-query, history
Pages (distilled knowledge) page, promote, refresh-page, inspect-page
Review governance review, review-queue, inspect-candidate, resolve-contradiction
Import / scan import-seed, import-local, scan
Health source-status, health, db-stats, rebuild-fts
Maintenance backup, benchmark, export

Full help: waypath --help.

Architecture

Waypath is built from four independent kernels behind a thin facade:

flowchart TD
    subgraph HOST[" Host Shims "]
        direction LR
        CX["codex"]
        CC["claude-code"]
        MC["mcp-server"]
    end

    Facade["<b>Facade</b><br/><code>createFacade()</code>"]

    TK["<b>Truth Kernel</b><br/>decisions · entities · preferences<br/>temporal validity · supersede"]
    AK["<b>Archive Kernel</b><br/>evidence · content-hash dedup<br/>FTS5 index"]
    ON["<b>Ontology</b><br/>graph traversal<br/>pattern expansion"]
    PR["<b>Promotion Engine</b><br/>candidate review<br/>contradiction detection"]

    HOST --> Facade
    Facade --> TK
    Facade --> AK
    Facade --> ON
    Facade --> PR

    classDef kernel fill:#21262d,color:#c9d1d9,stroke:#30363d,stroke-width:1px
    classDef facade fill:#1f6feb,color:#ffffff,stroke:#58a6ff,stroke-width:2px
    classDef host fill:#161b22,color:#c9d1d9,stroke:#30363d,stroke-width:1px
    class TK,AK,ON,PR kernel
    class Facade facade
    class CX,CC,MC host
  • Truth kernel — canonical decisions, entities, preferences, temporal validity (schema v3 with supersede + history).
  • Archive kernel — raw evidence store with content-hash dedup and FTS5 full-text index.
  • Ontology layer — graph traversal for entity/decision context expansion (patterns: project_context, person_context, system_reasoning, contradiction_lookup).
  • Promotion engine — candidate review, contradiction detection, supersede flows.

A single createFacade() exposes 14 verbs. Host shims adapt it to each agent's bootstrap protocol.

Configuration

Waypath is zero-config by default. To tune retrieval weights, adapter toggles, or review thresholds, drop a config.toml in your working directory (or point WAYPATH_CONFIG_PATH at one):

[source_adapters]
jarvis-memory-db = true
jarvis-brain-db  = false

[retrieval.source_system_weights]
truth-kernel = 1.2

[retrieval.source_kind_weights]
decision = 0.9
memory   = 0.5

[review_queue]
limit = 12

Override anything via env vars:

export WAYPATH_RECALL_WEIGHT_SOURCE_SYSTEM_TRUTH_KERNEL=1.8
export WAYPATH_REVIEW_QUEUE_LIMIT=8

Priority: env override > config.toml > built-in defaults.

MCP server

Waypath ships a native MCP (Model Context Protocol) server as a second binary:

waypath-mcp-server

Or via the main CLI:

waypath mcp-server --store-path ~/.waypath/project.db

Tools exposed via MCP: recall, page, promote, review, graph-query, source-status.

Requirements

  • Node.js ≥ 22.0 (required)
  • Node.js ≥ 22.5 recommended — unlocks native node:sqlite
  • better-sqlite3 is an optional fallback auto-used on 22.0–22.4 or where native sqlite is unavailable

Status

  • Version: 0.1.0 — first public release
  • Tests: 131 passing (unit + integration + benchmark)
  • Stable surface: CLI (26 commands), MCP server, facade API
  • Deferred: hosted deployment, multi-user sync, adaptive ranking feedback

Compared to alternatives

Waypath Cloud memory (mem0, zep) Vector-only RAG
Local-first ✓ ✗ depends
Canonical truth schema ✓ ✗ ✗
Graph-aware recall ✓ partial ✗
Explicit review gate ✓ ✗ ✗
MCP server built-in ✓ ✗ ✗
One-file install ✓ needs service varies

Contributing

Waypath welcomes host shims, source adapters, and bug fixes. Good first issues are labeled accordingly.

Read CONTRIBUTING.md for dev setup, code style, and PR flow.

Before submitting a PR:

npm run build
npm test

License

MIT © TheStack.ai — see LICENSE.

README copy captured when this server was indexed; it may lag the repository. The original is authoritative.

Questions

About TheStack-ai/waypath

How do I install TheStack-ai/waypath?

No package has been published for this server, so install it from source: clone the repository and follow its README.

Is TheStack-ai/waypath safe to use with an AI agent?

Its trust score is 34 out of 100 (low). It passes 1 of the 1 static security checks run so far. It has no ConduID identity yet, so agent calls to it are not receipted.

Is TheStack-ai/waypath still maintained?

Conduid hasn't recorded a commit date for this repository yet. Check the repository directly for recent activity.