Contextix

A CLI toolkit for agentic AI. Turn any MCP server — or RSS, markdown, URLs — into a typed knowledge graph. Query it from your terminal, your agent, or over MCP.

# Any MCP server → graph. One command.
npx contextix ingest mcp ./hackernews-top.mjs

# Or plain sources
npx contextix ingest markdown ~/notes
npx contextix ingest rss https://www.coindesk.com/arc/outboundfeeds/rss/

# Query the graph
npx contextix why "AI export controls"
npx contextix connect "Federal Reserve" "Bitcoin"

No Python. No Docker. No Neo4j. One npx command, your data stays local at ~/.contextix/graph.json.

Why

Your agent is smart about code, dumb about the world. Document RAG gives it text; vector search gives it snippets. Neither tells it what happened, who's involved, and how it's connected.

Contextix builds a typed causal graph from sources you choose. Agents query it via CLI — the same way they already call git, rg, or curl — so it slots into Claude Code, Cursor, Codex, Aider, or any shell-capable agent. MCP mode is bundled for Claude Desktop and MCP-native clients.

The wedge: the MCP ecosystem has 10,000+ servers exposing structured data — CoinGecko, arXiv, HackerNews, GitHub, Notion, Linear, your own — and zero graph layer on top. Contextix is that layer.

Install

npm install -g contextix
contextix --help

# or one-shot
npx contextix ingest markdown ~/notes

Requires Node 20+. Optional: ANTHROPIC_API_KEY env var for agentic extraction (falls back to regex mode).

CLI

Ingest — point at sources

contextix ingest mcp <skill-file>       # Any MCP server via a skill file (JS/TS)
contextix ingest rss <url>              # RSS / Atom / RDF feed
contextix ingest markdown <dir>         # Markdown vault (frontmatter + [[wikilinks]])
contextix ingest url <url>              # Single page, OG/Twitter meta + body
contextix ingest json <file|dir>        # Pre-formatted graph fragment

Each ingest run:

1. Fetches / reads source (or runs a skill against an MCP server)
2. Runs extraction — agentic (Haiku 4.5 with tool-use) when ANTHROPIC_API_KEY is set, regex otherwise. MCP skills bypass the extractor since they produce structured output directly.
3. Dedups entities (BTC / Bitcoin / bitcoin → one canonical node)
4. Merges into ~/.contextix/graph.json with valid_from timestamps

Force a specific mode with --extractor agentic|regex|auto or env CONTEXTIX_EXTRACTOR.

MCP ingest specifics

A skill is a single .mjs or .js file that tells contextix how to talk to one MCP server and what to extract. Skills live anywhere — drop one next to your project, commit to a repo, or put it in ~/.contextix/skills/.

# Keyless: Hacker News top stories → 20 events + author entities
contextix ingest mcp ./examples/skills/hackernews-top.mjs

# With env var: CoinGecko market snapshot
COINGECKO_DEMO_API_KEY=CG-xxx \
  contextix ingest mcp ./examples/skills/coingecko-markets.mjs

# Recent arXiv AI papers with author entities
contextix ingest mcp ./examples/skills/arxiv-ai.mjs

Skill file anatomy (JS, exports one defineSkill object):

import { defineSkill } from "contextix/skill";

export default defineSkill({
  name: "coingecko-markets",
  description: "Top 20 coins + global market snapshot",
  mcpServer: {
    command: "npx",
    args: ["-y", "@coingecko/coingecko-mcp"],
    env: { COINGECKO_DEMO_API_KEY: "${COINGECKO_DEMO_API_KEY}" },
  },
  requiredEnv: ["COINGECKO_DEMO_API_KEY"],
  async run({ mcp, emit, log }) {
    const result = await mcp.callTool({ name: "get_coins_markets", arguments: { vs_currency: "usd", per_page: 20 } });
    const coins = JSON.parse(result.content[0].text);
    for (const c of coins) {
      emit.entity({ entityType: "token", name: c.symbol.toUpperCase(), aliases: [c.name], domain: "crypto" });
      emit.event({ title: `${c.name} 24h: ${c.price_change_percentage_24h}%`, sourceName: "CoinGecko", importance: "low", tags: ["market-data"] });
    }
    log(`emitted ${coins.length}`);
  },
});

Full skill reference: examples/skills/README.md.

Markdown ingest specifics

• Walks recursively, skips .git, node_modules, .obsidian, .trash, _templates
• Parses YAML frontmatter (date, domain, tags) — flat key/value + inline/block lists
• Wikilinks [[X]] become concept entities with related_to edges from the note
• File mtime is used as detectedAt when frontmatter lacks date

Query — ask the graph

contextix signals                       # Recent events (24h default)
contextix signals --domain crypto -t 7d
contextix why "<event>"                 # Causal chain (BFS backward)
contextix connect "<a>" "<b>"           # Shortest path between entities
contextix entities --search "fed"       # Entity lookup

Output is human-readable by default. --json for piping.

contextix signals --json | jq '.events[] | select(.importance == "CRITICAL")'

Serve — MCP mode

contextix serve                         # stdio MCP server (default)

Same graph, exposed as 5 MCP tools: contextix_signals, contextix_why, contextix_connect, contextix_entities, contextix_graph.

Export

contextix export --format json          # Full graph dump
contextix export --format mermaid       # Mermaid diagram (roadmap)
contextix export --format cypher        # Cypher for Neo4j import (roadmap)

Use with your agent

Claude Code (bash tool)

Your agent calls contextix directly — no MCP needed:

"ingest my daily reads then tell me why the market moved"

Claude Code runs:

contextix ingest rss https://feeds.bloomberg.com/markets/news.rss
contextix why "S&P drop" --depth 3

Claude Desktop / Cursor (MCP)

.mcp.json:

{
  "mcpServers": {
    "contextix": {
      "command": "npx",
      "args": ["contextix", "serve"]
    }
  }
}

Scripts / cron

# Nightly ingest
0 2 * * * contextix ingest rss https://example.com/feed.xml

Data model

SignalEvent ──causes──▶ SignalEvent
     │                        │
  involves                involves
     ▼                        ▼
  Entity  ──influences──▶  Entity

• Edge types: causes, caused_by, correlates, involves, influences, precedes, contradicts
• Bi-temporal: every edge has valid_from / valid_until; invalidated edges are kept so you can reconstruct the graph at any point in time
• Confidence: every edge carries a [0,1] score + an evidence string
• Entity resolution: fuzzy dedup via string similarity; canonical node + alias list
• Storage: one JSON file at ~/.contextix/graph.json. Portable, inspectable, git-friendly

Full schema: src/graph/types.ts.

What contextix is not

| | Contextix | GraphRAG / LightRAG | mcp-memory | Graphiti | |---|---|---|---|---| | Install | npx contextix | pip + indexing | MCP only | pip + Neo4j | | MCP ecosystem ingest | ✅ via skills | ❌ | ❌ | ❌ | | File / feed ingest | ✅ RSS / md / URL | ❌ docs only | ❌ manual writes | ❌ SDK calls | | CLI interface | ✅ primary | ❌ Python scripts | ❌ | ❌ Python | | MCP server mode | ✅ bundled | ❌ | ✅ only | ✅ | | Local file graph | ✅ graph.json | ❌ | ✅ jsonl | ❌ Neo4j |

Contextix is not a RAG system, not a vector database, not a memory store for conversations. It's an agentic CLI that turns MCP servers, feeds, and files into a queryable typed graph.

Dogfood: contextix.io

contextix.io runs contextix on live crypto and AI sources. See what the graph looks like in production before you run it yourself.

Roadmap

See ROADMAP.md. Top priorities:

1. More skills — ship reference skills for Notion, Linear, GitHub, Slack, Gmail, Polymarket
2. Skill distribution — contextix skills install @contextix/crypto-pack style registry
3. Bring-your-own-model — OpenAI, Ollama, local LLM support (currently Claude Haiku 4.5 + regex)
4. Graph query depth — PageRank, temporal decay, contradiction detection
5. Hosted graph — optional --hosted mode pulls curated crypto/AI graph from contextix.io

Contributing

See CONTRIBUTING.md. Highest-impact areas:

• New skills — one .mjs file per MCP server. See examples/skills. No compile step, easy to write.
• New connectors — non-MCP source types (Slack export, Readwise, RSS variants). Each one is a function in src/ingest/.
• Extraction prompts — improve agentic entity/relation extraction for the non-MCP paths
• Query algorithms — src/graph/query.ts (PageRank, confidence propagation, temporal decay)
• Seed graph — verified events in data/seed-graph.json

Star the repo if this is the graph tool you wanted to exist. File issues for MCP servers you'd want a skill for.

License

MIT.