prolog-reasoner

SWI-Prolog as a "logic calculator" for LLMs — available as an MCP server and a Python library. Eliminate the black box from LLM logical reasoning.

LLMs excel at natural language but struggle with formal logic. Prolog excels at logical reasoning but can't process natural language. prolog-reasoner bridges this gap by exposing SWI-Prolog execution to LLMs.

Does it help?

On the built-in 30-problem logic benchmark:

Pipeline	Accuracy
LLM-only (claude-sonnet-4-6)	22/30 (73.3%)
LLM + prolog-reasoner	27/30 (90.0%)

The gap concentrates in constraint satisfaction and multi-step reasoning — the combinatorial territory LLMs are weak on and Prolog is strong on. Full breakdown below.

Why it works

LLMs pattern-match; Prolog actually searches and solves. When the LLM writes its problem down as Prolog, two things happen at once:

• Prolog handles the combinatorial work LLMs are weak on — constraint satisfaction, multi-step inference, exhaustive search.
• The reasoning exists as code you can read, re-run, and debug. When it goes wrong, you see the exact Prolog that failed and why.

Two ways to use it

• MCP server — Claude (or any MCP client) calls it as a logic solver during conversation. Rule bases let the LLM save stable domain rules once and reference them by name per call.
• Python library — full NL→Prolog pipeline with self-correction. Requires OpenAI or Anthropic.

Features

• MCP tools: execute_prolog for arbitrary SWI-Prolog execution, plus list_rule_bases / get_rule_base / save_rule_base / delete_rule_base for reusable named rule bases (v14)
• Rule bases: save stable Prolog rules once (e.g. chess move rules, legal axioms) and reference them by name from execute_prolog so the LLM only writes the situation-specific facts per call
• Transparent intermediate representation: the Prolog code is the audit trail — inspect, modify, or verify before execution
• CLP(FD) support: constraint logic programming for scheduling and optimization
• Negation-as-failure, recursion, all standard SWI-Prolog features
• Library mode: NL→Prolog translation with self-correction loop (OpenAI / Anthropic)

Requirements

• Python ≥ 3.10
• SWI-Prolog installed and on PATH (≥ 9.0)
• API key for OpenAI or Anthropic — only for library mode, not for the MCP server

Installation

# MCP server only (no LLM dependencies)
pip install prolog-reasoner

# Library with OpenAI
pip install prolog-reasoner[openai]

# Library with Anthropic
pip install prolog-reasoner[anthropic]

# Both providers
pip install prolog-reasoner[all]

MCP Server Setup

The MCP server exposes five tools — execute_prolog runs Prolog code written by the connected LLM, and four rule-base tools manage named, reusable Prolog modules. It does not call any external LLM API, so no API key is required.

Claude Desktop / Claude Code

{
  "mcpServers": {
    "prolog-reasoner": {
      "command": "uvx",
      "args": ["prolog-reasoner"]
    }
  }
}

Or, if prolog-reasoner is installed directly:

{
  "mcpServers": {
    "prolog-reasoner": {
      "command": "prolog-reasoner"
    }
  }
}

Docker (SWI-Prolog bundled)

Use Docker if you don't want to install SWI-Prolog locally:

docker build -f docker/Dockerfile -t prolog-reasoner .

{
  "mcpServers": {
    "prolog-reasoner": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "prolog-reasoner"]
    }
  }
}

Tool reference

execute_prolog(prolog_code, query, rule_bases=None, max_results=100, trace=False)

• prolog_code — Prolog facts and rules (string)
• query — Prolog query to run, e.g. "mortal(X)" (string)
• rule_bases — optional list of saved rule base names to prepend to prolog_code (in order). Use this to reuse stable domain rules across calls without re-sending them
• max_results — cap the number of solutions returned (default 100)
• trace — when True, attach a structured proof tree per solution to metadata.proof_trace. Opt-in sub-feature; has performance overhead and does not support CLP(FD), higher-order predicates, or assert/retract.

Returns a JSON object with success, output, query, error, and metadata.

On success, metadata includes execution_time_ms, result_count, truncated, and rule_bases_used. When rule bases were requested, rule_base_load_ms is also attached (disk I/O timing). On failure, metadata also includes error_category (one of syntax_error, undefined_predicate, unbound_variable, type_error, domain_error, evaluation_error, permission_error, timeout, trace_mechanism_error, unknown) and error_explanation — a natural-language hint for the connected LLM (or human) to decide how to fix the Prolog code.

Rule base tools — manage named, reusable Prolog modules under PROLOG_REASONER_RULES_DIR (defaults to ~/.prolog-reasoner/rules/). Names are restricted to [a-z0-9_-], length 1–64.

• save_rule_base(name, content) — write or overwrite a rule base. Content is syntax-validated (parse-only) before the write; failures surface as RULEBASE_003. Returns {"success": true, "name": ..., "created": bool} where created is true on first write, false on overwrite. Files over max_rule_size are rejected with RULEBASE_005.
• list_rule_bases() — return all saved rule bases with name, description, and tags. Metadata is extracted from leading % description: / % tags: comments in each file.
• get_rule_base(name) — return the raw Prolog source of a saved rule base.
• delete_rule_base(name) — remove a saved rule base.

For name/size/existence errors, the tools return {"success": false, "error": "...", "error_code": "RULEBASE_001"|"RULEBASE_002"|"RULEBASE_003"|"RULEBASE_005"} rather than raising. I/O failures (RULEBASE_004) are propagated as infrastructure errors.

Rule base conventions — start each rule base file with leading comments that double as list_rule_bases metadata:

% description: Chess piece movement rules
% tags: chess, games

piece_move(knight, (X1,Y1), (X2,Y2)) :- ...

Then reference from execute_prolog:

{
  "rule_bases": ["chess_moves"],
  "prolog_code": "position(knight, (4,4)).",
  "query": "piece_move(knight, (4,4), Target)"
}

Rule bases also serve as the foundation for domain-specialized forks: ship a curated set (legal axioms, game rules, tax scenarios, etc.) bundled via BUNDLED_RULES_DIR as a ready-to-use reasoning package.

Library Usage

The library exposes PrologExecutor (Prolog-only, no LLM) and PrologReasoner (NL→Prolog pipeline, needs an LLM API key).

Execute Prolog directly (no LLM)

import asyncio
from prolog_reasoner.config import Settings
from prolog_reasoner.executor import PrologExecutor

async def main():
    settings = Settings()  # no API key needed
    executor = PrologExecutor(settings)
    result = await executor.execute(
        prolog_code="human(socrates). mortal(X) :- human(X).",
        query="mortal(X)",
    )
    print(result.output)  # mortal(socrates)

asyncio.run(main())

Full NL→Prolog pipeline (requires LLM API key)

import asyncio
from prolog_reasoner import PrologReasoner, TranslationRequest, ExecutionRequest
from prolog_reasoner.config import Settings
from prolog_reasoner.executor import PrologExecutor
from prolog_reasoner.translator import PrologTranslator
from prolog_reasoner.llm_client import LLMClient

async def main():
    settings = Settings(llm_api_key="sk-...")  # from env or explicit
    llm = LLMClient(
        provider=settings.llm_provider,
        api_key=settings.llm_api_key,
        model=settings.llm_model,
        timeout_seconds=settings.llm_timeout_seconds,
    )
    reasoner = PrologReasoner(
        translator=PrologTranslator(llm, settings),
        executor=PrologExecutor(settings),
    )
    translation = await reasoner.translate(
        TranslationRequest(query="Socrates is human. All humans are mortal. Is Socrates mortal?")
    )
    print(translation.prolog_code)
    result = await reasoner.execute(
        ExecutionRequest(prolog_code=translation.prolog_code, query=translation.suggested_query)
    )
    print(result.output)

asyncio.run(main())

Configuration

All settings via environment variables (prefix PROLOG_REASONER_):

Variable	Default	Required for
LLM_PROVIDER	openai	library (openai or anthropic)
LLM_API_KEY	""	library only — leave unset for MCP
LLM_MODEL	gpt-5.4-mini	library
LLM_TEMPERATURE	0.0	library
LLM_TIMEOUT_SECONDS	30.0	library
SWIPL_PATH	swipl	both
EXECUTION_TIMEOUT_SECONDS	10.0	both
RULES_DIR	~/.prolog-reasoner/rules	both (where user-saved rule bases live)
BUNDLED_RULES_DIR	unset	both (optional — synced into RULES_DIR on first startup for shipping default rules with a fork)
MAX_RULE_SIZE	1048576 (1 MiB)	both (per-file save cap; save_rule_base rejects larger content with RULEBASE_005)
MAX_RULE_PROMPT_BYTES	65536 (64 KiB)	library only (total budget for the "Available rule bases" prompt section; truncated with a marker when exceeded)
LOG_LEVEL	INFO	both

Benchmark

benchmarks/ contains 30 logic problems across 5 categories (deduction, transitive, constraint, contradiction, multi-step) to compare LLM-only reasoning vs LLM+Prolog reasoning. The benchmark exercises the library path (translator + executor), since it requires the NL→Prolog step.

Results

Measured on anthropic/claude-sonnet-4-6, single run over 30 problems:

Pipeline	Accuracy	Avg latency
LLM-only	22/30 (73.3%)	1.7s
LLM + Prolog	27/30 (90.0%)	3.8s

Per-category breakdown:

Category	LLM-only	LLM + Prolog
deduction	6/6	6/6
transitive	6/6	5/6
constraint	3/7	6/7
contradiction	4/4	3/4
multi-step	3/7	7/7

The gap is concentrated in constraint (SEND+MORE, 6-queens, knapsack, K4 coloring, Einstein-lite) and multi-step (Nim game theory, 3-person knights-and-knaves, TSP-4, zebra puzzle) — exactly the combinatorial/search-heavy territory where symbolic solvers outperform pattern completion. On purely deductive or transitive questions the LLM is already strong and Prolog adds latency without accuracy gains.

All 3 LLM+Prolog failures were Prolog execution errors from malformed LLM-generated code (missing predicate definitions, unbound CLP(FD) variables) rather than reasoning errors — addressable via prompt tuning. Notably, every failure is inspectable: you can see the exact Prolog that failed and why, rather than a wrong natural-language answer with no explanation.

Running it yourself

docker run --rm -e PROLOG_REASONER_LLM_API_KEY=sk-... \
    prolog-reasoner-dev python benchmarks/run_benchmark.py

Results are saved to benchmarks/results.json.

Comparison with other Prolog MCPs

Several Prolog MCP servers exist, each with different design choices. prolog-reasoner is intentionally stateless and spot-use — Prolog is a calculator you call when logic matters, not the backbone of your agent's memory.

	prolog-reasoner	Stateful Prolog MCPs
Prolog's role	Per-call reasoning tool	Project-wide knowledge base
State	Stateless execution (each call independent); optional named rule bases for reusable static rules, no inter-call session memory	Persistent sessions / layered KBs
Reproducibility	Same input (incl. same rule bases) → same output, always	Depends on accumulated state
Integration effort	Use where logic matters, skip where it doesn't	Architectural commitment
A/B testable vs LLM-only	Yes (each call is a controlled experiment)	Structurally not comparable

This is also why accuracy benchmarks are published here and not elsewhere: statelessness is what makes a side-by-side comparison possible.

If you need persistent agent memory, hallucination-safeguarded fact storage, or a full neuro-symbolic substrate, other projects may fit better:

• adamrybinski/prolog-mcp — Trealla WASM with save/load sessions
• umuro/prolog-mcp — layered KB with file-backed persistence
• vpursuit/model-context-lab — SWI-Prolog with security sandboxing
• dr3d/prolog-reasoning — neuro-symbolic memory with write-path safety

We're the spot-use option.

Development

# Build dev image
docker build -f docker/Dockerfile -t prolog-reasoner-dev .

# Run tests (no API key needed — LLM calls are mocked)
docker run --rm prolog-reasoner-dev

# With coverage
docker run --rm prolog-reasoner-dev pytest tests/ -v --cov=prolog_reasoner

# Or via docker compose
docker compose -f docker/docker-compose.yml run --rm test

License

MIT