Euclid-MCP

MCP server for logical reasoning — turns facts into formal proofs.

Euclid-MCP is a hybrid cognitive architecture: a lightweight LLM describes the world in facts, and a deterministic engine performs the actual deduction. The LLM never needs to reason — it only needs to describe.

With Euclid-MCP, an 8B model can solve reasoning tasks that stump even 400B+ cloud models — because the engine handles deduction deterministically. Every answer comes with a proof tree, so you can trace why a conclusion holds, not just what it is. Use it to enforce RBAC policies, audit cloud compliance, validate loan eligibility rules, or reason over any domain where answers must be explainable and verifiable.

Euclid-MCP is written in Python and uses Euclid-IR, a human-readable intermediate language designed for both AI agents and humans. It currently uses SWI-Prolog as its inference engine and can be consumed in multiple ways: via MCP by AI agents (OpenCode, Claude, Cursor), via HTTP by tools and automation platforms (n8n, Zapier, Make), and via Python API for direct integration. Euclid-IR rules can also be used to augment RAG pipelines with deterministic policy enforcement.

How it works

┌──────────────┐     ┌──────────────────┐     ┌──────────────┐     ┌─────────────────┐
│  LLM/Agent   │────▶│  Euclid-MCP      │────▶│  Translator  │────▶│  SWI-Prolog     │
│  (MCP Client)│◀────│  (MCPServer)     │◀────│  + Meta-IP   │◀────│ (persistent)    │
└──────────────┘     └──────────────────┘     └──────────────┘     └─────────────────┘

1. Receive facts, rules, and a query in a simple intermediate language
2. Translate into Prolog with a meta-interpreter for proof tree capture
3. Execute via a persistent SWI-Prolog engine process (JSON-lines protocol on stdin/stdout; the workspace is reloaded per call, no process spawn overhead)
4. Return solutions + proof trees as structured JSON

Additional tools (explain, diagnose, what_if, check_kb) extend this core flow with natural-language explanations, analysis, scenario testing, and validation.

LLMs describe. Euclid MCP proves.

Knowledge Base

For small knowledge bases, facts and rules can be provided with each request.

Since v0.2.0 a knowledge base can be loaded at server startup and reused across calls, so agents only pass the session-specific facts for the current query.

This minimizes token usage, improves performance, and allows small LLMs to reason over large rule sets without reconstructing the entire knowledge base for every request.

Scalability

The engine is persistent since v0.3.0 — a single long-lived SWI-Prolog process per server instance, reloaded per request over a JSON-lines pipe instead of booting Prolog for every call. Requests stay stateless: each one brings its own knowledge base (or uses the preloaded one), so instances share nothing.

This makes Euclid-MCP horizontally scalable:

• HTTP API — run any number of instances behind a load balancer (nginx, a Kubernetes Service, …). No session affinity needed: any instance can serve any request.
• MCP stdio — each MCP client spawns its own instance by design, giving natural isolation and parallelism across clients.
• Resource footprint — one swipl process per instance (~tens of MB) instead of one short-lived process per request, so a single instance serves many requests cheaply.

A single instance handles one request at a time; an in-process engine pool for concurrent requests is on the roadmap (see IDEAS.md).

Intermediate Language

Even if currently Euclid-MCP uses a Prolog Engine, no Prolog syntax is required.
Euclid-IR (Intermediate Representation) is a declarative intermediate representation for logical inference. Variables use $name, implication is IF, conjunction is AND.

Text format:

human(socrates)
mortal($x) IF human($x)

? mortal($who)

YAML format:

facts:
  - parent(tom, bob)
  - parent(bob, ann)
  - parent(tom, liz)
rules:
  - ancestor($x, $y) IF parent($x, $y)
  - ancestor($x, $y) IF parent($x, $z) AND ancestor($z, $y)

query: ancestor(tom, $who)

Full language reference: docs/EUCLID_IR.md

Euclid-IR Syntax Reference

Element	Syntax	Example
Facts	predicate(args)	parent(tom, bob)
Variables	$name (lowercase)	$who, $x, $count
Implication	IF	mortal($x) IF human($x)
Conjunction	AND	p($x) AND q($x)
Negation	NOT	NOT active($user)
Query	? predicate	? ancestor(tom, $who)
String literals	"..." or '...'	"alice@example.com"
Multi-line rules	Body on next line	rule($x) IF\n body($x)

Arithmetic Comparisons

Rules support arithmetic comparisons that are evaluated during deduction:

# Stale access: users who haven't logged in for 90+ days
stale_access($user) IF
    user($user) AND last_login_days($user, $days) AND $days > 90

# Excessive permissions: more than 15 direct permissions
excessive_permissions($user, $count) IF
    user($user) AND permission_count($user, $count) AND $count > 15

# Clearance check: user clearance >= resource classification
can_access($user, $resource) IF
    user($user) AND resource($resource, _, _, _, _, $cls) AND
    classification($cls, $cls_level, _) AND
    user_clearance($user, $user_level) AND $user_level >= $cls_level

Supported operators: >, >=, <, <=, ==, is, !=

Multi-line Rules

Rules can span multiple lines for readability:

can_deploy($user, $env) IF
    user($user) AND
    has_role($user, $role) AND
    deploy_requires_level($env, $min) AND
    deploy_role_level($role, $level) AND
    $level >= $min AND
    user_has_permission($user, deploy_code)

Conjunctions in Queries

Queries can combine multiple predicates:

? can_access_resource($who, $res) AND resource($res, _, _, _, _, secret)

This returns solutions where both conditions are satisfied simultaneously.

Why External Inference?

The external inference gives several advantages:

• deterministic
• explainable
• verifiable
• inexpensive
• replaceable backend

In the current implementation Euclid-MCP uses Prolog.
Prolog is a 50-year-old battle-tested logic engine. Using it as a "deduction coprocessor" lets small LLMs perform complex multi-step reasoning without needing larger, more expensive models. The intermediate language strips away Prolog's syntax quirks while keeping its logical core.

Some internal benchmarks demonstrate the difference: with 1 000+ facts, LLMs alone score 2/5 while Euclid-MCP scores 5/5 — and runs 7× faster while outputting 14× fewer tokens.

Tools

Euclid-MCP exposes 5 tools, each with a specific purpose:

Tool	Purpose
reason	Main deduction — get solutions + proof trees
explain	Readable, natural-language reasoning steps
diagnose	Understand why a query succeeds or fails
what_if	Test modifications before applying them
check_kb	Validate KB consistency before reasoning

reason

Main tool for verifiable deterministic reasoning.

Parameter	Type	Default	Description
knowledge	string	—	Facts & rules in text or YAML format
query	string?	—	Override query (optional)
max_solutions	int	5	Max solutions to return
max_depth	int	30	Max proof tree depth

Returns ReasonResult with solutions[] — each containing variable bindings and a proof tree.

explain

Deterministic proof-tree → natural-language reasoning steps. No LLM involved: it walks the proof tree of each solution and renders every step in plain language, citing the rule ID (# RULE: <id>) when a rule has one. Use it to turn a proof into an auditable, human-readable explanation.

Parameter	Type	Default	Description
knowledge	string?	—	Facts & rules in text or YAML format
query	string?	—	Override query (optional)
max_solutions	int	5	Max solutions to return
max_depth	int	30	Max proof tree depth

Returns ExplanationResult with explanations[] — each containing variable bindings and an ordered list of natural-language steps.

diagnose

Query analysis — understand why a query succeeds or fails.

Parameter	Type	Default	Description
knowledge	string	—	Facts & rules in text or YAML format
query	string	—	Query to diagnose
mode	string	why	One of: why, why_not, what_needs
max_solutions	int	5	Max solutions to return
max_depth	int	30	Max proof tree depth

Modes:

• why — explain why a query holds (or that it doesn't)
• why_not — explain why a query fails (missing facts/rules)
• what_needs — suggest what would make a false query true

Returns DiagnosisResult with holds, findings[], conclusion, and optionally proof.

what_if

Scenario analysis — apply modifications to a knowledge base and compare results.

Parameter	Type	Default	Description
base_knowledge	string	—	Base facts & rules
modifications	string	—	+ fact(...) to add, - fact(...) to remove
query	string	—	Query to evaluate
max_solutions	int	5	Max solutions to return
max_depth	int	30	Max proof tree depth

Returns WhatIfResult with before_count, after_count, delta, solutions_before, solutions_after, conclusion.

check_kb

Knowledge base validator — check for consistency before running deduction.

Parameter	Type	Default	Description
knowledge	string	—	Facts & rules in text or YAML format

Returns KBCheckResult with valid, errors[], warnings[], facts_count, rules_count, predicates_count.

KB Preload (v0.2.0)

Since v0.2.0 a knowledge base can be loaded once at server startup and reused across calls, so agents only pass the session-specific facts for the current query.

Preload a KB by file path, via the EUCLID_KB_PATH environment variable or a --kb-path CLI flag:

# Environment variable
EUCLID_KB_PATH=/path/to/policies.euclid python3 -m euclid_mcp

# CLI flag (MCP stdio, console script, and HTTP API)
python3 -m euclid_mcp --kb-path /path/to/policies.euclid
python3 integrations/euclid_api.py --kb-path /path/to/policies.euclid --port 8080

Behavior:

• The file is validated with check_kb at startup and the server fails fast with a clear message if the file is missing, unreadable, oversized, or invalid.
• knowledge/base_knowledge on reason, explain, diagnose, what_if, and check_kb become optional: an explicit value always wins, an empty value falls back to the preloaded KB. With neither, tools return a clear "No knowledge provided" error.
• A markdown digest of the preloaded KB (fact/rule/predicate counts, predicate inventory, rules with their IDs) is appended to the server instructions, so agents can see what the KB covers without extra tool calls.

Backward compatible: passing knowledge explicitly behaves exactly as before.

Installation

pip

# Prerequisites: Python ≥ 3.10, SWI-Prolog
brew install swi-prolog

# Install
pip install euclid-mcp

From source

git clone https://github.com/meob/Euclid-MCP
cd Euclid-MCP
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Docker

No local SWI-Prolog installation needed — the image bundles everything.

# Build
docker build -t euclid-mcp .

# MCP stdio mode (for local MCP clients)
docker compose run --rm euclid-mcp

# HTTP API mode (for n8n, Zapier, remote access)
docker compose up euclid-api
# API available at http://localhost:8080

See Docker in Integrations for full details.

Usage

Via MCP (OpenCode, Claude, etc.)

{
  "mcpServers": {
    "euclid-mcp": {
      "command": "python3",
      "args": ["-m", "euclid_mcp"],
      "cwd": "/path/to/euclid-mcp"
    }
  }
}

Via Python

from euclid_mcp.server import reason, explain, diagnose, what_if, check_kb

# Reasoning
result = reason(knowledge="""
    human(socrates)
    mortal($x) IF human($x)
    ? mortal($who)
""")
for sol in result.solutions:
    print(sol.substitutions, sol.proof.type)

# Explanation — readable reasoning steps (cites rule IDs when present)
expl = explain(
    knowledge="human(socrates)\nmortal($x) IF human($x)  # RULE: BIO-001",
    query="mortal($who)"
)
for e in expl.explanations:
    print(e.substitutions, e.steps)

# Diagnosis — why does a query fail?
diag = diagnose(
    knowledge="human(socrates)\nmortal($x) IF human($x)",
    query="mortal(plato)",
    mode="why_not"
)
print(diag.conclusion)

# What-if — how does adding a fact change results?
scenario = what_if(
    base_knowledge="human(socrates)\nmortal($x) IF human($x)",
    modifications="+ human(plato)",
    query="mortal($who)"
)
print(f"Before: {scenario.before_count}, After: {scenario.after_count}")

# KB validation
check = check_kb(knowledge="human(socrates)\nmortal($x) IF human($x)")
print(f"Valid: {check.valid}, Errors: {check.errors}")

Example output

{
  "query": "ancestor(tom, $who)",
  "solutions": [
    {
      "substitutions": {"who": "bob"},
      "proof": {
        "type": "rule",
        "goal": "ancestor(tom, bob)",
        "body": "parent(tom, bob)",
        "rule_id": "GEN-1",
        "subproof": {"type": "fact", "goal": "parent(tom, bob)"}
      }
    },
    {
      "substitutions": {"who": "ann"},
      "proof": {
        "type": "rule",
        "goal": "ancestor(tom, ann)",
        "body": "parent(tom, bob), ancestor(bob, ann)",
        "rule_id": "GEN-2",
        "subproof": {
          "type": "and",
          "left": {"type": "fact", "goal": "parent(tom, bob)"},
          "right": {
            "type": "rule",
            "goal": "ancestor(bob, ann)",
            "body": "parent(bob, ann)",
            "rule_id": "GEN-1",
            "subproof": {"type": "fact", "goal": "parent(bob, ann)"}
          }
        }
      }
    }
  ]
}

Rules can carry an audit-trail ID via a trailing # RULE: <id> comment; the ID is surfaced as rule_id on the rule nodes of the proof tree, so a decision can be cited ("this derives from rule GEN-2").

Diagnose output

{
  "query": "mortal(plato)",
  "mode": "why_not",
  "holds": false,
  "findings": [
    {
      "type": "satisfied",
      "predicate": "human",
      "detail": "Facts exist for 'human' (1 facts)"
    }
  ],
  "conclusion": "The query fails. Check rule conditions."
}

What-if output

{
  "query": "mortal($who)",
  "modifications": "+ human(plato)",
  "before_count": 1,
  "after_count": 2,
  "delta": "more",
  "solutions_before": [{"substitutions": {"who": "socrates"}}],
  "solutions_after": [
    {"substitutions": {"who": "plato"}},
    {"substitutions": {"who": "socrates"}}
  ],
  "conclusion": "Solutions increased: 1 -> 2."
}

Explain output

{
  "query": "mortal($who)",
  "explanations": [
    {
      "substitutions": {"who": "socrates"},
      "steps": [
        "mortal(socrates) is derived by rule BIO-001 from: human(socrates).",
        "human(socrates) is asserted as a fact in the knowledge base."
      ]
    }
  ]
}

Use cases

• Small LLM reasoning: Offload deduction from LLMs (3-8B) to a deterministic engine
• Explainable decisions: Every answer comes with a proof tree which allows explanation, reasoning trace, and justification
• Business rules: Validate logic chains (permissions, workflows, compliance)
• Dependency analysis: Circular dependency detection, topological ordering
• Education: Interactive logic tutoring with visible proof chains
• Knowledge preload: Complex business rules can be loaded in Euclid instead of using a vector database
• Query diagnosis: Understand why queries fail and what facts/rules are missing
• Scenario analysis: Test "what-if" modifications before applying them to production
• KB validation: Check knowledge bases for consistency before reasoning

Real-world examples

There are several examples provided as samples: Genealogy, RBAC, Classification, Loan Eligibility, Cluedo Detective, IT Security & Compliance, LLM vs Euclid-MCP, ... Most interesting ones are the IT Security & Compliance (with CIS, AWS, IAM Standards enforcement, Company Policies implementation, hundreds of Data Facts) and side-by-side LLM vs Euclid-MCP.

Examples full description: docs/EXAMPLES.md

Integrations

OpenCode

Euclid-MCP includes a pre-configured agent in .opencode.json:

{
  "mcpServers": {
    "euclid-mcp": {
      "command": "python3",
      "args": ["-m", "euclid_mcp"],
      "cwd": "."
    }
  },
  "agents": {
    "reasoning-engine": {
      "description": "Deterministic logic engine",
      "instructions": "Write facts in Euclid IR, use the reason tool...",
      "mcpServers": ["euclid-mcp"]
    }
  }
}

n8n / Zapier / Make

Run the HTTP API:

python3 integrations/euclid_api.py --port 8080

Endpoint	Method	Purpose
/reason	POST	Deduction with proof trees
/explain	POST	Natural-language reasoning steps
/diagnose	POST	Query failure analysis
/what-if	POST	Scenario testing
/check-kb	POST	KB validation
/health	GET	Health check

# Reasoning
curl -X POST http://localhost:8080/reason \
  -H "Content-Type: application/json" \
  -d '{"knowledge": "human(socrates)\nmortal($x) IF human($x)\n? mortal($who)"}'

# Explanation
curl -X POST http://localhost:8080/explain \
  -H "Content-Type: application/json" \
  -d '{"knowledge": "human(socrates)\nmortal($x) IF human($x)\n? mortal($who)"}'

# Diagnosis
curl -X POST http://localhost:8080/diagnose \
  -H "Content-Type: application/json" \
  -d '{"knowledge": "human(socrates)\nmortal($x) IF human($x)", "query": "mortal(plato)", "mode": "why_not"}'

# What-if
curl -X POST http://localhost:8080/what-if \
  -H "Content-Type: application/json" \
  -d '{"base_knowledge": "human(socrates)\nmortal($x) IF human($x)", "modifications": "+ human(plato)", "query": "mortal($who)"}'

# KB validation
curl -X POST http://localhost:8080/check-kb \
  -H "Content-Type: application/json" \
  -d '{"knowledge": "human(socrates)\nmortal($x) IF human($x)"}'

Docker

The Docker image bundles SWI-Prolog + Python, so no local prerequisites are needed. Base image: swipl:stable (Debian Bookworm).

Two modes via docker-compose:

# MCP stdio — pipe to a local MCP client
docker compose run --rm euclid-mcp

# HTTP API — expose REST endpoints on port 8080
docker compose up euclid-api

Standalone usage:

# Build
docker build -t euclid-mcp .

# Run HTTP API
docker run --rm -p 8080:8080 euclid-mcp \
  python3 integrations/euclid_api.py --port 8080

# Run MCP stdio (interactive)
docker run --rm -i euclid-mcp

# Quick test — reason directly from CLI
docker run --rm euclid-mcp python3 -c "
from euclid_mcp.server import reason
r = reason(knowledge='human(socrates)\nmortal(\$x) IF human(\$x)\n? mortal(\$who)')
print(r.solutions[0].substitutions)
"

Docker image size: ~370 MB (SWI-Prolog + Python 3.11 + dependencies).

CLI pipeline

echo '{"knowledge": "red(apple)\\n? red($x)"}' | python3 integrations/euclid_cli.py

See integrations/README.md for full details.

Development

Requirements: Python ≥ 3.10, SWI-Prolog.

# Install in editable mode with dev dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Lint
ruff check .

# Type check
mypy euclid_mcp integrations

# Tests with coverage
pytest --cov=euclid_mcp --cov=integrations

The CI workflow (.github/workflows/ci.yml) runs these same checks on push and pull request, across Python 3.10–3.12.

Logging & tracing

Every tool call is logged with its name, elapsed time, and outcome. Enable structured logs by setting EUCLID_LOG_LEVEL (one of DEBUG, INFO, WARNING, ERROR, CRITICAL) — e.g. EUCLID_LOG_LEVEL=INFO. Without the variable, only warnings and errors are emitted.

The HTTP API also supports request tracing: send an X-Request-Id header and it is echoed back on the response and included in the access logs.

How is Euclid?

Euclid was an ancient Greek mathematician. Living and teaching in Alexandria, he built the foundations of geometry and number theory using rigorous logical proofs.

Euclid-MCP is not:

• an LLM
• a knowledge base
• a vector database
• an agent framework
• a planner

Euclid-MCP is a deterministic inference engine that can be used by any of them.
Euclid-MCP allows deterministic and explainable replies from small LLMs on Edge hardware too.

License

Apache 2.0