Predicate
An MCP server that provides a self-improving knowledge graph with per-triple provenance and deterministic reasoning, enabling auditable, reproducible, and contradiction-aware answers for AI agents.
README
<div align="center">
Predicate
Reasoning memory for AI agents — a self-improving knowledge graph that grows with use.
Every fact carries its provenance. Every answer carries its derivation. The schema sharpens itself the more your agent uses it.
Install · How it works · Architecture · Packages
</div>
What is Predicate?
Most agent "memory" is a flat pile of text: the model re-reads it and takes it on faith — no record of where a fact came from, no guarantee the same question answers the same way twice, and no signal when two notes contradict each other. It never gets smarter; it just gets bigger.
Predicate is different. It stores facts as RDF triples with per-triple provenance and confidence, and answers questions through a deterministic reasoner that produces an explanation path for every derived claim. And it learns its own shape: the schema is versioned like code and evolves under a propose → validate → use-gated promotion loop, so the structure your agent relies on gets sharper the more it's used — and prunes what it doesn't.
The bet is not that a graph recalls facts better than a long context window —
at the scales we've measured (predicate-eval/SCALE-FINDINGS.md),
in-context recall is competitive and often simpler. The bet is that a graph makes
answers auditable, reproducible, and contradiction-aware in a way a flat pile
can't.
Why it's different
- Auditable answers.
kg_explainreturns the chain of triples and rules that produced a claim — each step cited back to its source and confidence. Not "trust me," but "here's the derivation." - Contradictions surface instead of averaging out. When two sources disagree about a fact the schema marks functional or disjoint, the reasoner flags the conflict rather than silently picking one.
- The graph remembers, and the schema earns its keep. Facts persist across sessions. New schema becomes durable only after real queries reference it within a TTL — unused proposals expire on their own, so the graph can't thrash.
- Goal-conditioned growth. Concepts enter the graph because a goal needed them, not because a document mentioned them.
Install / use it in your agent
Prerequisite: Node 20+. Nothing else — no Docker, no database to run.
The fastest path is the predicate-skill npm package, which bundles the MCP
server, the predicate CLI, and per-client integration adapters.
<details open> <summary><strong>Claude Code</strong> — one-command marketplace install</summary>
/plugin marketplace add NordicAgents/predicate
/plugin install predicate@predicate
Restart Claude Code (or /reload-plugins), then:
predicate up # creates the local store + the 8 named graphs
predicate doctor # all checks green
Hooks, slash commands, and the kg_* tools are wired automatically. Slash
commands: /predicate:up, /predicate:down, /predicate:status,
/predicate:ask <question>.
Marketplace install failing? Register Predicate as a plain MCP server instead — see Manual install below. The reasoning tools work; only automatic capture and slash commands are lost.
</details>
<details> <summary><strong>Codex CLI</strong> (native plugin, capture supported)</summary>
codex plugin marketplace add NordicAgents/predicate
# then enable "predicate" in the interactive plugin browser
Set [features] plugin_hooks = true in ~/.codex/config.toml and approve the
hooks once via /hooks. See packages/predicate-skill/hooks/codex-cli/README.md.
Plugin install failing? Register Predicate as a plain MCP server instead — see Manual install below (reasoning tools work; no capture).
</details>
<details> <summary><strong>VS Code Copilot</strong> / <strong>Cursor</strong> (MCP-only)</summary>
From your project root:
npx predicate-skill install vscode # writes .vscode/mcp.json + AGENTS.md
npx predicate-skill install cursor # writes .cursor/mcp.json + AGENTS.md
Restart the editor. Reasoning tools work; there is no automatic turn capture (neither host exposes usable lifecycle hooks).
</details>
Manual install (fallback for any host)
If a native plugin/marketplace install fails — or your host isn't listed above
(Continue.dev, OpenCode, any stdio MCP client) — run Predicate as a plain stdio
MCP server. You get all 10 kg_* reasoning tools on any MCP-capable host; the
only thing you lose is automatic Stop-hook capture (and, on Claude Code, the
slash commands).
npm install -g predicate-skill # published package; ships the server + CLI
predicate up # create the local store + named graphs
predicate doctor # all checks green
# The MCP server command is (copy the resolved absolute path):
node "$(npm root -g)/predicate-skill/server.bundle.mjs"
No env vars are required: the server defaults to the disk-backed Oxigraph store
(PREDICATE_BACKEND=oxigraph, PREDICATE_DATASET=predicate) — identical to the
native installs. Set them only to override.
Register that command with your host:
Claude Code
claude mcp add predicate -- node "$(npm root -g)/predicate-skill/server.bundle.mjs"
Codex CLI — add to ~/.codex/config.toml (use the absolute path printed by npm root -g):
[mcp_servers.predicate]
command = "node"
args = ["/ABSOLUTE/PATH/predicate-skill/server.bundle.mjs"]
Continue.dev — in ~/.continue/config.yaml:
mcpServers:
- name: predicate
command: node
args: ["/ABSOLUTE/PATH/predicate-skill/server.bundle.mjs"]
Full per-client matrix, the 10 MCP tools, the CLI reference, and config live in the package README:
packages/predicate-skill/README.md.
Architecture
%%{init: {
"theme": "base",
"themeVariables": {
"fontSize": "15px",
"primaryColor": "#FFD23F",
"primaryTextColor": "#1A1A1A",
"primaryBorderColor": "#1A1A1A",
"lineColor": "#1A1A1A",
"edgeLabelBackground": "#FFF7E1"
}
}}%%
flowchart LR
Agent["<b>your agent</b><br/>Claude Code · Codex"]:::actor
Hook["<b>Stop hook</b><br/>turn capture"]:::actor
subgraph Predicate["PREDICATE"]
direction TB
Tools["<b>10 kg_* tools</b><br/>stdio MCP"]:::iface
Reasoner["<b>21-rule reasoner</b><br/>16 OWL 2 RL + 5 domain<br/>+ SHACL"]:::compute
Storage["<b>Storage adapter</b><br/>Oxigraph · Fuseki"]:::data
Graphs[("<b>8 named graphs</b><br/>tbox · tbox-staging · abox<br/>inferred · provenance · meta<br/>goals · usage")]:::store
end
Agent -->|MCP| Tools
Hook -->|assert| Tools
Tools -->|SPARQL| Storage
Storage <--> Graphs
Reasoner -.->|CONSTRUCT fixpoint| Graphs
Tools -.-> Reasoner
classDef actor fill:#FFD23F,stroke:#1A1A1A,stroke-width:3px,color:#1A1A1A
classDef iface fill:#4ECDC4,stroke:#1A1A1A,stroke-width:3px,color:#1A1A1A
classDef compute fill:#FF6B6B,stroke:#1A1A1A,stroke-width:3px,color:#FFFFFF
classDef data fill:#1A535C,stroke:#1A1A1A,stroke-width:3px,color:#FFFFFF
classDef store fill:#FFFFFF,stroke:#1A1A1A,stroke-width:3px,color:#1A1A1A
style Predicate fill:#FFF7E1,stroke:#1A1A1A,stroke-width:3px,color:#1A1A1A
The agent reads the schema (kg_explore_schema), drafts SPARQL, asserts facts
with provenance (kg_assert), and asks questions (kg_ask). The reasoner does
the logic; the model formulates queries and interprets results.
Runtime flow
%%{init: {
"theme": "base",
"themeVariables": {
"fontSize": "14px",
"actorBkg": "#FFD23F",
"actorBorder": "#1A1A1A",
"actorTextColor": "#1A1A1A",
"actorLineColor": "#1A1A1A",
"signalColor": "#1A1A1A",
"signalTextColor": "#1A1A1A",
"noteBkgColor": "#FF6B6B",
"noteTextColor": "#FFFFFF",
"noteBorderColor": "#1A1A1A",
"labelBoxBkgColor": "#4ECDC4",
"labelBoxBorderColor": "#1A1A1A",
"labelTextColor": "#1A1A1A",
"sequenceNumberColor": "#FFFFFF",
"activationBkgColor": "#FFD23F",
"activationBorderColor": "#1A1A1A"
}
}}%%
sequenceDiagram
participant Host as Agent host
participant Stop as stop.sh hook
participant Ext as turn-extractor.ts
participant KG as kg_* tools
participant R as Reasoner
participant G as Named graphs
rect rgb(255, 247, 225)
Note over Host,G: TURN ENDS — capture
Host->>Stop: turn payload (stdin)
Stop->>Ext: predicate extract --from-stdin
Ext->>KG: kg_assert (typed triples + provenance)
KG->>G: write kg:abox + kg:provenance
R->>G: forward-chain to fixpoint → kg:inferred
end
rect rgb(230, 250, 248)
Note over Host,G: NEXT TURN — ask
Host->>KG: kg_explore_schema / kg_ask
KG->>G: SPARQL
G-->>KG: results
Host->>KG: kg_explain (why?)
KG-->>Host: derivation path (cited triples + rules)
end
More diagrams (schema lifecycle, scale findings) live in
docs/diagrams/.
How it works
- Storage. 8 named graphs separate slow-changing schema (
kg:tbox) from fast-flowing facts (kg:abox), materialized entailments (kg:inferred), per-triple metadata (kg:provenance), goals (kg:goals), usage logs (kg:usage), staging (kg:tbox-staging), and version history (kg:meta). - Reasoning. A curated set of 21 rules — 16 OWL 2 RL rules plus 5 domain
and judgment derivations (
Hotspot,FlakyCommand,ActiveFile, current-judgment, conflict detection). The entailment rules run as SPARQLCONSTRUCTforward-chained to a fixpoint; disjointness is checked as an inconsistency pass; SHACL shapes add closed-world validation. The OWL 2 RL coverage is a scoped subset — in v1,owl:propertyChainAxiomis limited to length-2 chains andowl:hasKeyto single-property keys. - Provenance. Every triple is annotated with source, time, confidence, and extraction method using RDF-star. Low-confidence triples stay visible to queries but are excluded from the inference closure, so they cannot poison entailment.
- Schema lifecycle. The agent proposes deltas to
kg:tbox-staging, never tokg:tboxdirectly. A proposal is promoted only after the reasoner accepts it and it has been referenced by N successful queries inside a TTL; unused proposals expire quietly. - Cross-session continuity. A Stop hook extracts typed triples from each turn
— files modified, commands that passed or failed — into
kg:abox. The reasoner derivesHotspot,FlakyCommand, andActiveFileso the next session can ask "what's unstable here?" without re-reading the repo.
Packages
This is a pnpm monorepo. Each package has its own README with details.
| Package | Purpose |
|---|---|
predicate-mcp |
MCP server, the 10 kg_* tools, and the storage adapters (Oxigraph + Fuseki). |
predicate-reasoner |
21-rule forward-chaining reasoner + SHACL validation + inference traces for kg_explain. |
predicate-agent |
Goal store, decomposer, gap detector, schema proposer, promotion sweeper, generalizer, lifecycle controller. |
predicate-cli |
The predicate command-line interface. |
predicate-ontology |
Versioned TBox catalog, SHACL shapes, meta vocabulary. |
predicate-server |
Backend bootstrap + Fuseki/TDB2 docker-compose for the opt-in backend. |
predicate-eval |
End-to-end demo + multi-hop eval + ontology CI harness. |
predicate-skill |
The distributable npm package — bundled server + CLI + SKILL.md + per-client hooks. |
Development
git clone https://github.com/NordicAgents/predicate
cd predicate
pnpm install
pnpm build # builds all packages + the plugin bundle
pnpm test # runs against the default Oxigraph backend, no Docker needed
# for the Fuseki leg: PREDICATE_BACKEND=fuseki + a running Fuseki
Other root scripts: pnpm typecheck, pnpm lint, pnpm fuseki:up,
pnpm fuseki:down.
License
Elastic License 2.0 (ELv2) — source-available. See
LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.