Tugra

Provenance-aware memory format for AI agents: every claim carries its source, age and boundary.

AI & MLTypeScriptv0.1.4

tugra

npx tugra init

That creates a vault, writes a sample fact, and prints a config block. Paste the block into your MCP client. On a TTY, npx tugra prints help and exits. Piped (Claude Desktop, Cursor, Claude Code) it is the MCP server.

Provenance-aware memory for AI agents. Every claim carries its source, its age, and its boundary. There is no cloud.

What it is

A fact that cannot name where it came from is not a fact. Tugra stores each claim as a file whose frontmatter holds source, last verification date, shelf life, and — when the topic is off-limits — a boundary that forbids invention. Search ranks by token score, then freshness, then confidence. Retired and rotten facts stay out of the default set.

Tools

ToolWhat it does
fact_searchSearch the vault. Retired/rotten omitted unless archive: true.
fact_readRead one fact by uid. Body is escaped before the model sees it.
fact_proposeWrite a draft. Secret patterns are rejected before any write. type: "boundary" is always quarantined.
event_reportAppend a local telemetry line. No network.

Stored field names stay in the vault's native shape (kaynak, guven, raf_omru, sinir). The tool names and parameter names above are the public contract.

Install — env paths (optional)

tugra init is enough to start. Override the two paths only if you already have a vault elsewhere. Without them, the server looks next to the installed package — that is wrong for a bare npx with no init.

  • TUGRA_VAULT — vault (markdown facts)
  • TUGRA_EVENTS — telemetry directory

Authorization: if no authorization store is configured, single-user mode is on — search and propose work without a profile. If an authorization store is configured (a yetki/ directory, or TUGRA_AUTH), each agent needs a JSON profile or search returns unauthorized.

Claude Desktop

claude_desktop_config.json:

{
  "mcpServers": {
    "tugra": {
      "command": "npx",
      "args": ["-y", "tugra"],
      "env": {
        "TUGRA_VAULT": "/absolute/path/to/vault",
        "TUGRA_EVENTS": "/absolute/path/to/events"
      }
    }
  }
}

Claude Code

.mcp.json at the project root, or claude mcp add:

{
  "mcpServers": {
    "tugra": {
      "command": "npx",
      "args": ["-y", "tugra"],
      "env": {
        "TUGRA_VAULT": "/absolute/path/to/vault",
        "TUGRA_EVENTS": "/absolute/path/to/events"
      }
    }
  }
}

Cursor

.cursor/mcp.json or Cursor Settings → MCP:

{
  "mcpServers": {
    "tugra": {
      "command": "npx",
      "args": ["-y", "tugra"],
      "env": {
        "TUGRA_VAULT": "/absolute/path/to/vault",
        "TUGRA_EVENTS": "/absolute/path/to/events"
      }
    }
  }
}

Windsurf

mcp_config.json:

{
  "mcpServers": {
    "tugra": {
      "command": "npx",
      "args": ["-y", "tugra"],
      "env": {
        "TUGRA_VAULT": "/absolute/path/to/vault",
        "TUGRA_EVENTS": "/absolute/path/to/events"
      }
    }
  }
}

Codex

~/.codex/config.toml:

[mcp_servers.tugra]
command = "npx"
args = ["-y", "tugra"]

[mcp_servers.tugra.env]
TUGRA_VAULT = "/absolute/path/to/vault"
TUGRA_EVENTS = "/absolute/path/to/events"

Windows: use a full path (C:\\Users\\…\\vault). Node 20 or newer.

More client notes: docs/install.md.

Shared-vault authorization (optional)

Single-user setups do not need this. Add TUGRA_AUTH only when several agents share one vault and each needs its own profile (mcp-readonly@tugra and others as JSON files in that directory). A missing profile then returns unauthorized. An empty TUGRA_AUTH is treated as unset — single-user mode stays on.

Host library surface (not the MCP wire)

The MCP tools (fact_search, fact_read, fact_propose, event_report) enforce authorization on every call. The published package also ships dist-paket/akis.js and dist-paket/yetki.js so a host application (cron, mirror, cockpit) can write telemetry without going through JSON-RPC.

Those modules are public on purpose. akisBildir({ atlaYetki: true }), eylem: "yetki_talebi", and dosyaYoksaIzin (default true) skip or relax the check. harcamaEkle mutates a profile. The host that imports them owns authorization. The MCP wire cannot set these flags — the tool schema does not accept them.

Scale vault vs target vault

A0–A5 levels are facts (yonetisim.yetki.a0 … a5) in a vault. The stdio server reads them from TUGRA_VAULT, or from the cockpit kasa/ when that variable is unset.

tugraArac / createTugraMcp take an optional kasaKok (the write/search target). Scale does not follow that target. It defaults to varsayilanKasa() — the same central vault the stdio server uses. A host that points kasaKok at a data-only tree keeps using the cockpit / TUGRA_VAULT scale. To read scale from a different tree, pass skalaKasa explicitly.

All four tools share one resolver. This is the contract: separate target + central governance stays reachable. YAYIN/12 briefly defaulted scale to kasaKok; that broke the split-root host. YAYIN/13 restores the central default.

What we do not guarantee

  • No cloud sync. The vault is the files you pointed at. Nothing is uploaded.
  • No automatic merge. Two writers, two files. You reconcile.
  • No delete in this release. Retirement exists; erasure is later.
  • No automatic conflict detection. Contradictory facts can sit side by side until a human says otherwise.
  • No hosted service. npx tugra is a local stdio process.

This package is not published as a SaaS. There is no price table here.

Requirements

  • Node.js 20 or newer. This is a support decision, not a technical floor: the package is tested on 20 and 22 in CI, and it also runs on 18 — but 18 is past its end of life, so we do not support it.
  • A vault directory you own

Topic map (optional)

<vault>/_konu-haritasi.json — { "desen", "bayrak", "konu" } rules in harita, alt_kirilim, and stem. If the file is missing the map is empty: unknown text falls back to kurum.genel or dunya.<world>.genel. Broken or over-long patterns are skipped and logged. This package does not ship a company taxonomy.

License

Apache-2.0. See LICENSE and NOTICE.

The marketing page lives in ../site/ (npm run preview there). It is not deployed from this package.

Compatibility: TUGRA_KASA, TUGRA_AKIS, TUGRA_YETKI (and the older TALAMUS_* / MULTI_* names) still work as a fallback when the English name is unset.

Installation

Source-derived launch command. Check the maintainer’s required arguments and credentials before running:

bash
npx -y tugra

Set up in your AI client

Merge this template into ~/Library/Application Support/Claude/claude_desktop_config.json. Keep existing servers. Add any arguments, credentials, and permissions required by the maintainer; this template has not been install-tested.

json
{
  "mcpServers": {
    "com-tugra-ai-tugra": {
      "command": "npx",
      "args": [
        "-y",
        "tugra"
      ]
    }
  }
}

Restart Claude Desktop completely for changes to take effect. Confirm the server appears connected in the client’s tool list, then try a read-only example from its documentation.

Claude Desktop setup reference

Package

tugranpm

Compatible MCP Clients

Tugra works with any MCP-compatible client. Copy the config snippet from the Configuration section above and add it to the file shown for your client, then restart the application.

  • Claude Desktop~/Library/Application Support/Claude/claude_desktop_config.jsonRestart Claude Desktop completely for changes to take effect.
  • Cursor~/.cursor/mcp.jsonRestart Cursor for changes to take effect.
  • VS Code.vscode/mcp.jsonReload VS Code window for changes to take effect.
  • Windsurf~/.codeium/windsurf/mcp_config.jsonRestart Windsurf for changes to take effect.
  • Claude Code.mcp.jsonSave at the project root, then start Claude Code in that project and review the MCP server approval prompt. Keep real credentials out of shared files.

Learn More