Living Atlas MCP
Enables local tools to interact with a private-first knowledge graph, with full CRUD access to authorized data while remote AI providers only see approved plaintext through policy-scoped MCP surfaces.
README
Living Atlas
Living Atlas is a private-first knowledge graph system for a Logseq/Obsidian-inspired graph. It stores complete graph bytes in a Cloudflare deployment for anywhere access, keeps sensitive plaintext available only to local/keyholding clients, and exposes policy-scoped MCP surfaces so local tools can see the full authorized graph while remote AI providers only see approved remote-readable data.
System Layers
Living Atlas has two tightly linked halves:
- Runtime architecture: storage, Cloudflare/local materialization, encryption, sync, MCP access, CRUD, audit, conflict handling, and privacy boundaries.
- Knowledge semantics: nodes, edges, predicates, events, bitemporal dates, relationship vocabulary, and Logseq/Obsidian migration semantics.
They are one product. If the two tracks appear to conflict, runtime/security architecture controls where data lives and who can access it; temporal-edge docs control what graph facts mean.
Document Map
- PRD - product goals, users, requirements, non-goals.
- Architecture Requirements - system shape and constraints.
- V1 Architecture Decisions - accepted V1 runtime and privacy decisions.
- Knowledge Schema Runtime Integration - how temporal-edge semantics map onto the runtime/storage architecture.
- ADR 0001 - Cloudflare byte custody with local-key sensitive sync.
- ADR 0002 - separate local and remote MCP authority.
- ADR 0003 - visible, auditable CRUD history.
- Runtime Topology Options - chosen V1 topology plus alternate deployment profiles.
- Cloudflare-First Bootstrap And Local Sync - first deployment, safe authority claim, browser-keyed setup, local link, and sync.
- Public Repo And Personal Cloudflare Deployment - public template/repo boundaries, Terraform/Wrangler split, and private personal deployment state.
- Remote-Primary With Local Sensitive Path - remote MCP for normal work, local/keyholding path for sensitive plaintext, and future federation hooks.
- Local MCP Boundary - local MCP as private authority and release producer, not a remote-call backend.
- Complete Cloudflare Custody Diagram - Cloudflare stores the complete graph while sensitive content remains local-key-only.
- Live Graph Activity And Audit - near-live graph firing view plus repeatable CRUD audit/replay.
- 100M Scale Plan - segmented storage, indexes, compaction, and sync design for large graphs.
- Offline Sync And Conflict Resolution - continuous sync, offline queues, generations, and conflict handling.
- Key Management - KEK/DEK hierarchy, envelopes, device enrollment, revocation, and release keys.
- Access Modes - remote-safe, cloud-unlock session, and local-keyholding security modes.
- Identity, Configuration, And Key Control Plane - user/device/client setup, capability grants, key config, recovery, and admin surfaces.
- Event Subsystems - sync change log, durable audit ledger, and live activity stream.
- Metadata Leakage Budget - Cloudflare-visible metadata and path/index constraints.
- Compaction And Retention - tombstones, snapshots, long-offline clients, and erasure.
- Local MCP Authentication - local auth, capabilities, admin mode, and localhost threat model.
- Security and Access Model - trust tiers, encryption, policy enforcement.
- CRUD Observability - how create/read/update/delete activity is seen and audited.
- Implementation Plan - build phases and validation gates.
- Development Readiness Checklist - first build slice, pre-deploy gates, and before-real-data tests.
- Private Cloudflare Deployment Overlay - recommended private repo pattern for account-specific Cloudflare deployment state.
- Temporal Edge Model - schema package entrypoint for edge/event ontology and migration semantics.
- Contributing, Security, and Code of Conduct - public collaboration and reporting policies.
Working Thesis
Living Atlas is not a single hosted plaintext brain. It is one knowledge graph with Cloudflare complete custody, a local complete replica, access-classed objects, and separate capability surfaces:
- Cloudflare custody: complete graph bytes, including sensitive ciphertext.
- Remote MCP: Cloudflare-hosted CRUD for remote-readable data.
- Local replica: complete graph bytes plus local decrypted/indexed views.
- Local MCP: full authorized graph CRUD with local keys.
- Cloud-unlock session: optional remote convenience mode where a transient key may unlock sensitive content in the cloud runtime, without key persistence, when runtime-memory exposure is acceptable.
- Sensitive objects: plaintext CRUD only through keyholding client/local path.
- Praxis-facing contracts: headless activity, audit, replay, and graph APIs that a UI or operator runtime can consume without moving UI code into Atlas.
Current Status
Phase 1 scaffold exists as a TypeScript workspace. It includes contracts, synthetic fixtures, access policy evaluation, metadata leakage scanning, readiness check commands, Cloudflare first-claim bootstrap, fixture-backed local MCP tools, a sealed local keyring, an encrypted local graph store, sync batch persistence for remote-readable plaintext and sensitive ciphertext envelopes, envelope pull/replay, a token-gated remote MCP with remote-readable graph CRUD, edge CRUD, deterministic text search, bounded graph traversal, timeline queries, cloud-unlock decrypt for v1 AES-GCM inline ciphertext envelopes, and hash-only replay reporting over audit/activity/ operational events. It also includes a token-gated usage/budget endpoint that reports provider-neutral observed usage against configurable limits. It does not import real graph data or deploy personal Cloudflare resources.
Development
For the complete first-run runbook, see Development Readiness Checklist.
First-run synthetic sequence:
npx pnpm@11.8.0 install
npm run check
npm run smoke:local
npm run local:deploy-synthetic
npm run cloudflare:wrangler-smoke
npm run preflight:synthetic
These commands use synthetic fixtures and public-safe templates. They must not import real graph data, claim a real authority, publish personal Cloudflare values, or replace placeholder config with private deployment state.
Install with the pinned package manager:
npx pnpm@11.8.0 install
Run the local gate:
npm run check
Run the two synthetic local smoke flows:
npm run smoke:local
local:install-smoke exercises the local install mode: it creates a sealed
local control store and sealed local keyring, starts the local MCP over stdio,
calls the fixture graph read and synthetic CRUD tools, creates a local-private
plaintext draft through the MCP, and checks activity plus encrypted graph files
for token/sensitive-bait/plaintext leakage.
cloudflare:local-smoke exercises the Worker bootstrap and sync routes
in-process with fake D1/R2 bindings.
Run the full synthetic local deployment exercise before looking at Cloudflare:
npm run local:deploy-synthetic
This creates a temporary local profile, writes an encrypted local control store, starts the local MCP over stdio, performs synthetic read/create/update/ tombstone operations, boots the local Worker harness, claims bootstrap, pushes and pulls ciphertext sync batches through the sync daemon, checks stale and bad-token-binding rejection, and scans the resulting local artifacts for token/sensitive-bait leakage.
Run the local stress gate when changing CRUD, policy, sync, or leakage code:
npm run stress:local
Check deployed synthetic usage before running any live Cloudflare stress:
npm run cloudflare:live-usage-gate
npm run cloudflare:live-ops-report
npm run cloudflare:live-crud-tiny
This performs hundreds of synthetic local CRUD operations in one run, including duplicate creates, stale updates, invalid versions, empty patches, oversized objects, store-limit enforcement, tombstones, audit/activity checks, and leakage scans. It also pushes many one-generation ciphertext sync batches through the local Worker harness, verifies D1/R2 counts, pulls the batch summaries back, and checks malformed, stale, generation-gap, replay, bad-token, bad-binding, query-token, and invalid-pull behavior.
Run the full synthetic preflight before any Cloudflare deployment work:
npm run preflight:synthetic
This runs the repo gate, the full synthetic local deployment exercise, the local stress gate, the Wrangler dry-run smoke, Terraform/OpenTofu formatting, and Terraform/OpenTofu validation against public-safe example inputs.
npm run check runs the repo-safety/leakage check, TypeScript typecheck, and
Vitest suite. The check CLI's default all mode runs:
local: contract, policy, leakage, path opacity, and repo-safety checks.cloudflare-deploy-readiness: synthetic public-template deploy readiness, including placeholder Cloudflare bindings, no private deploy values, complete fixture manifest coverage, opaque paths, and no sensitive bait in Cloudflare metadata.first-run-guardrails: synthetic bootstrap checks for sealed/unclaimed first-run behavior, token-required claim, token burn, concurrent first-claim lock behavior, and token-in-query guard coverage.
Run individual checks while iterating:
npx tsx packages/check/src/cli.ts local
npx tsx packages/check/src/cli.ts cloudflare-deploy-readiness
npx tsx packages/check/src/cli.ts first-run-guardrails
npx tsx packages/check/src/cli.ts wrangler-local-runtime
npx pnpm@11.8.0 check works too and uses the same underlying gate.
Validate the Cloudflare infrastructure skeleton:
npm run infra:fmt
npm run infra:validate
Build the Worker example without deploying:
npx wrangler@4.103.0 deploy --dry-run \
--config packages/cloudflare-worker/wrangler.example.jsonc \
--outdir /tmp/living-atlas-worker-dry-run
This dry run is synthetic-only. It should validate the public Worker template, not claim an authority, upload real graph data, or publish personal Cloudflare account values. A real deployment uses a private/ignored overlay for the Cloudflare account, resource ids, deploy token, bootstrap claim-token hash, sync token hash, domains, and state.
Workspace packages:
@living-atlas/contracts: object envelopes, identity/config records, capability types, temporal edge/event validators, audit/change contracts, and operational observability events.@living-atlas/fixtures: synthetic-only fixture graph with sensitive bait and remote-safe content.@living-atlas/policy: capability-bound policy evaluator and output filters.@living-atlas/leakage: bait-string scanner, opaque Cloudflare path helpers, and public-repo safety scanner.@living-atlas/check: local scaffold verification CLI.@living-atlas/cloudflare-worker: Cloudflare Worker routes and Durable Object first-claim bootstrap lock skeleton, plus token-gated sync batch persistence/status, envelope pull through R2/D1 bindings, remote MCP remote-readable graph CRUD/search/traversal/timeline/edge tools, and redacted structured request observability.@living-atlas/local-control-store: encrypted local authority/control-plane state store, local profile path helpers, and fixture generation tooling.@living-atlas/local-keyring: sealed local keyring and AES-GCM payload encryption helpers for local install mode.@living-atlas/local-graph-store: durable snapshot/journal graph replica for local CRUD and sync replay, with redacted or local-keyring-encrypted persistence by policy.@living-atlas/local-mcp: local trusted-ingress MCP skeleton with bearer token capability checks, sealed control-store loading, fixture graph status/list/read plus synthetic CRUD tools backed by in-memory fixtures or the durable local graph store, and redacted audit events.@living-atlas/sync-agent: local sync-agent skeleton that builds ciphertext batches for sensitive local graph changes, tracks an in-memory synthetic outbox/daemon plan, submits to the Worker sync route, fetches remote summaries/envelopes, and applies pulled envelopes into the local graph store with version-conflict reporting.@living-atlas/atlas-client: dependency-light TypeScript client helpers for Praxis and other consumers calling remote MCP, activity, and usage surfaces with token headers and redacted error details.@living-atlas/activity-replay: hash-only replay inspection and reporting over durable audit, live activity, and operational observability events.@living-atlas/cloudflare-workeralso exposes/api/usage/statusfor health-token-gated observed usage and configurable budget ratios. The response shape is generic so non-Cloudflare deployments can implement the same contract with provider-specific collectors.- The Worker and remote MCP also expose a
living-atlas-usage-gate:v1safe-to-test/stop-testing decision. The gate is tunable per deployment and is intended to fail closed before live synthetic stress runs. cloudflare:live-ops-reportadds a compact operator report over the gate and provider-side inventory available through bound Cloudflare services, including R2 object count/byte reconciliation.
Launch the fixture local MCP server with generated synthetic control state:
LIVING_ATLAS_LOCAL_MCP_TOKEN='replace-with-local-dev-token' \
npm run local-mcp:fixture
Run it from an MCP client or the Inspector; a direct terminal run waits on stdio.
Create an encrypted synthetic local control store for local MCP development:
LIVING_ATLAS_LOCAL_CONTROL_STORE=/tmp/living-atlas-control-store.json \
LIVING_ATLAS_LOCAL_CONTROL_STORE_PASSPHRASE='replace-with-local-dev-passphrase' \
LIVING_ATLAS_LOCAL_MCP_TOKEN='replace-with-local-dev-token' \
npm run local-control:fixture-store
Then launch the fixture local MCP server from that sealed store:
LIVING_ATLAS_LOCAL_CONTROL_STORE=/tmp/living-atlas-control-store.json \
LIVING_ATLAS_LOCAL_CONTROL_STORE_PASSPHRASE='replace-with-local-dev-passphrase' \
npm run local-mcp:fixture
Cloudflare templates:
packages/cloudflare-worker/wrangler.example.jsonc: public-safe Worker config example with placeholder bindings and no bootstrap secret.infra/cloudflare/modules/living-atlas-single-authority: reusable Terraform/OpenTofu module for R2, D1, and KV resources.infra/cloudflare/examples/single-authority: public-safe validation example that expects the Cloudflare account id from private environment input.
Cloudflare is the complete graph byte custodian, not the plaintext authority. Sensitive/local-private graph content is stored in Cloudflare as ciphertext and opaque metadata; only local or browser keyholding clients decrypt it and build full private indexes. Remote MCP may serve explicitly remote-readable projections, but that is separate from the private graph ciphertext source of truth.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。