zeusi-mcp
OfficialGenerates embeddings for memory chunks through OpenAI, enabling semantic and hybrid search with a hard monthly spend cap.
Stores sourced team memory in PostgreSQL with pgvector, including schema migrations, row-level security, wall rules, least-privilege roles, and hybrid search over memory records.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Zeusi
Zeusi is a sourced memory for a team and its AI assistants. Standing rules, current focus, decisions, notes and work history live in Postgres with pgvector. Every row carries its source (source_uri, source_ts, a content hash and the ingest run that wrote it), so an assistant can cite where an answer came from and never has to invent one.
Per-user workspace encryption is enabled by setting ZEUSI_WORKSPACE_ENCRYPTION=on; new personal workspaces are then encrypted from sign-up, and you hold the key. On the hosted service at zeusi.ai: Your memory is encrypted. Only you hold the key. The workspace is not end-to-end encrypted: the server holds sealed data and unlocks it when you sign in.
What it gives you:
A schema for sourced memory: facts, decisions (with the exact words and who said them), episodes, notes and chunks, people, PRs and reviews, routine runs. Supersede functions instead of edits; views for the current state.
A wall in the database. A trigger on every content table runs the same verdict as the importers (
src/wall.ts==zeusi.wall_eval()): hard rules for areas, P&L paths and lanes, and protected names (people or companies whose content must never be stored, in any spelling, look-alike or percent-encoding), plus soft content rules that quarantine by uri and hash only. Row-level security hides thepersonalscope from every reader. The example nameplankivain this repo is fictional. Seedocs/wall.md.Per-user encryption when
ZEUSI_WORKSPACE_ENCRYPTION=on: new personal workspaces are encrypted from sign-up, and you hold the key.Least-privilege roles, created by SQL with no passwords: an ingest login, a backup login, read-only logins for agents, a narrow decision-write login. See
docs/roles.md.A vault importer for markdown notes (read-only on the vault, two wall gates, deletion guard, advisory lock), embeddings with a hard monthly spend cap, and hybrid search (vector + full text).
An MCP connector (
zeusi-mcp) that any MCP client can use:zeusi_startup,zeusi_search,zeusi_search_feedback,zeusi_context_for_task,zeusi_onboarding_statusand an opt-inzeusi_log_decision, with an output wall on every item. A canonical agent skill ships inskills/zeusi-startup.
What Zeusi is
Zeusi is a sourced memory for a team and its AI assistants. A Zeusi decision is context, not permission. Decisions and locks returned by Zeusi are cited context for the assistant; they are not an approval or a GO to act now. Anything that needs the owner's approval still needs the owner's current, explicit approval.
Related MCP server: Genomefy MCP Server
Quick start (local stdio MCP)
Requirements: Node.js 22.18+ and PostgreSQL 17+ with the vector extension (Neon works out of the box; any Postgres with pgvector works).
npm ci && npm run build && npm test # tests run on PGlite, no database neededDatabase and owner URL. Create a database. Put the owner connection strings in
.env.local(gitignored; names in.env.example): a pooledDATABASE_URLand a directDATABASE_URL_UNPOOLED.Migrate as the owner:
npm start -- migrate --dry-run, thennpm start -- migrate. On a database that already has migrations, take a snapshot first and pass--i-took-a-snapshot.Role passwords. The migrations create the login roles without passwords. Set them as the owner (
\password zeusi_box_ro, ...) and store each one in your password manager, then on the machine that uses it in the macOS Keychain (zeusi-neon-<role>) or~/.pgpass(0600). Seedocs/roles.md.Service file. Copy
config/pg_service.example.confto~/.pg_service.conf, fill in hosts and database name,chmod 600. Check withnpm start -- doctorandnpm start -- doctor --roles.Sources and protected names. Copy
config/sources.example.tomltoconfig/sources.toml(gitignored) and point it at your vault. Put the people or companies that must stay out under[wall] protected_names, then apply their hard rules as the owner:node dist/cli.js wall seed-sql | psql "<owner URL>". Never commit the names. The example nameplankivais already seeded.Ingest and embed.
npm start -- ingest vault --dry-run(counts only, no database), then--real. Embeddings need an OpenAI key in the macOS Keychain itemzeusi-openai-api-key:npm start -- embed --estimate,embed --dry-run, thenembed --real(docs/embeddings.md).Connect an assistant over stdio:
{ "mcpServers": { "zeusi": { "command": "node", "args": ["/path/to/zeusi/dist/mcp.js"] } } }or
claude mcp add zeusi -- node /path/to/zeusi/dist/mcp.js. Install the skill fromskills/zeusi-startup/. Details and client snippets:docs/mcp.md.
Credentials stay on the machine that runs the server (~/.pg_service.conf plus Keychain or ~/.pgpass); they never go into a client config, a skill, a prompt or this repo.
Hosted mode
The official hosted server is https://mcp.zeusi.dev/mcp. To run your own, use node dist/mcp.js --hosted. Each assistant gets its own bearer token (zeusi mcp-token mint|list|revoke). The server fails closed (HTTP 503 until a valid tokens file exists). See docs/hosting.md. Client snippets are in docs/mcp.md under "Startup snippets".
Site
The public site is live at https://zeusi.ai. It is deployed as static files from site/. See site/README.md.
Commands
npm start -- <command> (or node dist/cli.js <command>); --help lists every option.
Command | Connects as | What it does |
| role / none | TLS, channel binding, server major, pgvector, migrations; role catalog checks. No hosts or credentials printed. |
| none / | Import the configured vault notes through both wall gates. Dry run prints counts only. |
| none / | Offline cost estimate, or embed pending chunks under a hard monthly cap. |
|
| Semantic (with the key) or lexical search; every hit cites |
| role | Quarantined items (no content); rows that break the live wall (exit 1 if any). |
| none | Owner SQL for the protected names in the private config. |
| owner, direct | Apply |
| none | Offline token admin for the hosted MCP server. |
Development
npm run typecheck,npm run build,npm test(Node's test runner on PGlite; no network, no database).npm run hooks:installadds a pre-commit hook that runs gitleaks;npm run secret-scanscans the history.npm run checkruns typecheck, build, tests, a gitleaks history scan and the repo guards.
License
MIT. Copyright (c) 2026 Sovra GmbH.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
shared AI-context layer for teams — persistent memory your agents search and update over MCP
Team memory for AI coding agents and code reviews, reviewed by humans and served over MCP.
Cross-tool persistent memory and context for AI assistants over MCP.
Shared cross-LLM long-term memory over MCP: semantic recall, sessions, and media (pgvector).
Related MCP Servers
- FlicenseCqualityDmaintenanceProvides RAG (Retrieval-Augmented Generation) capabilities via Contextual AI, enabling query processing and context-aware responses with citations. Integrates with MCP clients like Cursor IDE and Claude Desktop.1-
- AlicenseNot gradedqualityCmaintenanceProvides a local context-memory layer for AI assistants, enabling retrieval-augmented queries, explanations, feedback, and status checks via MCP tools.1Apache 2.0
- FlicenseAqualityCmaintenanceEnables AI agents to persist, search, and query shared memories in PostgreSQL through hybrid retrieval, preset answers, and AI-to-AI consultation via MCP.9-
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to persist, retrieve, and share source-linked memory across sessions via MCP tools, backed by PostgreSQL with project/session management, decision tracking, keyword search, and cross-host context handoff.Apache 2.0