poe2-build-mcp
An MCP server for Path of Exile 2: a queryable game corpus plus Path-of-Building-faithful calculations, so an LLM can import your build, answer questions, and theorycraft against real numbers (not invented ones).
README
poe2-build-mcp
A Path of Exile 2 build assistant for LLMs. It pairs a queryable game corpus with a Path-of-Building–faithful calculation engine, so an AI can import your build, answer questions, and theorycraft against real, computed numbers — never invented ones.
Install it as a one-click .mcpb in Claude Desktop and ask things like:
- "Import this PoB code — where's my DPS bottlenecked, and what's the cheapest fix?"
- "Build me a level-90 Witchhunter Detonate Living and make sure it caps resistances."
- "What supports should Spark use, and how much does each one actually add?"
- "Is this build tanky enough for endgame bosses? What's my weakest defensive layer?"
Every DPS / EHP / resistance figure it reports is computed by a real headless Path of Building, not guessed. When PoB can't model something, it says so instead of making a number up.
Install (Claude Desktop)
- Download the
.mcpbfor your OS from the latest release. - Open it with Claude Desktop and install it when prompted (extensions are managed under Settings → Extensions).
- Start asking build questions. The first call boots the engine (a few seconds to load game data); everything after is fast.
The bundle ships the PoB engine, LuaJIT, the game-data corpus, and Python deps. You need
Python 3.11+ available on your system. The corpus and engine self-update from validated
GitHub releases into a writable user-data folder, so the data stays current without a reinstall
(server code updates land when you install a newer .mcpb). See PACKAGING.md.
Starting a session
Once the extension is enabled, just ask a Path of Exile 2 question — Claude will reach for the
tools. To kick off explicitly, pick the start_build_session prompt from Claude Desktop's
prompt menu (or a task-specific one: analyze_build, build_from_goal, audit_defenses). In
Claude Code, configure the server (see Build from source) and,
if you like, add a line to your CLAUDE.md such as "For Path of Exile 2 build questions, use the
poe2-build MCP tools." so it prefers them automatically.
What it can do
Import & analyze a build from a PoB share code, a pobb.in/pastebin link, or raw PoB XML — then read back computed DPS, EHP, resistances, life/ES, and more.
Create & tweak builds from scratch: set class/ascendancy, level, skills + supports, gear,
config, and the passive tree, validating every change on the engine. There's a greedy passive
optimizer and an A/B compare_to.
Look things up offline: items, skill/support gems, affixes, uniques, passives, ascendancies — a bundled SQLite/FTS corpus, no network needed.
Reason well, not just compute: build_advice and explain_mechanic provide durable PoE2
optimization principles and mechanics references (see
server/BUILD_ADVICE.md) so the assistant knows what to change; the
engine confirms the effect.
Stay current: live currency/unique prices, corpus freshness checks, and one-click self-update.
The toolset (42 MCP tools)
Build / compute — real Path of Building numbers:
import_build(source)— PoB share code, pobb.in/pastebin link, or raw XMLget_build_stats(keys?)— computed stats (DPS, EHP, resistances, life/ES/mana, …)get_build()/export_build()— full read-back / export as a PoB import codeget_defenses()— resists (+over-cap), EHP, and the active resistance-penalty assumptionset_class(class, ascendancy?)·set_level(level)·set_skill(skill)·set_config(…)equip_item(raw)·unequip_item(slot)·list_config_options(query?)evaluate_build(goals)— pass/fail against numeric goals ·compare_to(source)— A/B deltassolve_for(metric, target, lever)— root-find the modifier magnitude needed to hit a stat targetsearch_passives(query?, node_type?)/get_passive(node)alloc_passive(node)/dealloc_passive(node)— allocate/route by id or name, with deltasoptimize_passives(metric, points)— greedy point allocation (metric="balanced"raises DPS+EHP)engine_health()— engine + install diagnostics (liveness, LuaJIT/tree/data/server versions)
Corpus / knowledge — bundled SQLite + FTS, no engine needed:
search_items(query, item_class?)/get_item(name_or_id)find_skills(query?, gem_type?, tag?, color?)/get_gem(name_or_id)/find_supports_for(skill)build_advice(topic?)— evergreen build-optimization principlesexplain_mechanic(topic)— concise mechanics references (resistances, ailments, spirit, EHP, …)search_mods(query, item_tag?, mod_type?)/reverse_lookup(stat)search_uniques(query, item_type?)/get_unique(name)parse_item(text)— parse an item's text → affix tiers (T1=best) + open prefix/suffix slotslist_ascendancies(character?)/corpus_info()
Live ops & self-update — network:
get_prices(query, kind, league?)— poe2scout currency/unique prices ·list_price_leagues()get_meta_builds(league?)— live ascendancy popularity (poe.ninja; context, not a recommendation)check_for_updates()/apply_updates()— pull validated engine + corpus releasescheck_data_version()/update_corpus(rebuild_from_source?)
The connector also ships an LLM-facing operating guide (delivered via the MCP instructions
channel, from server/ASSISTANT_GUIDE.md) plus workflow prompts
(analyze_build, build_from_goal, audit_defenses) so the assistant uses the tools cohesively.
How it works
Two layers behind one Python server:
- Compute — a vendored PathOfBuilding-PoE2 fork run headless under LuaJIT as a persistent JSON-RPC subprocess (loads game data once, answers many calls cheaply). PoB is pinned to a release tag and bumped deliberately behind golden-build tests.
- Knowledge — a bundled, read-only SQLite/FTS corpus built offline from RePoE/poe2db.
Knowledge code never imports compute code; cross-layer orchestration lives in server/main.py.
See PLAN.md for the full design and CLAUDE.md for engineering conventions.
Build from source (development)
Prerequisites: Python 3.11+, uv (python -m pip install uv), and LuaJIT 2.1
(Windows via MSYS2: pacman -S mingw-w64-ucrt-x86_64-luajit; auto-detected at
C:\msys64\ucrt64\bin\luajit.exe, override with the POB_LUAJIT env var). Then clone the
git-ignored PoB working copy (pinned in pob/PINNED.md):
git clone --depth 1 --branch dev \
https://github.com/PathOfBuildingCommunity/PathOfBuilding-PoE2.git \
pob/PathOfBuilding-PoE2
Set up and verify:
uv sync # create venv, install deps
uv run python -m pipeline.build_corpus # build data/corpus.sqlite from RePoE (network)
uv run pytest # golden-build suite (engine + corpus)
uv run python scripts/smoke_mcp_client.py # full MCP protocol over stdio (all tool groups)
The scripts/smoke_*.py files cover each tool group individually; pytest is the pinned
golden-value regression suite (see tests/). On Windows, run the smoke scripts with
PYTHONUTF8=1 to avoid code-page issues with some item/skill names (the server itself is
unaffected). Lint/type with uv run ruff check . && uv run mypy server.
To run from source in Claude Desktop, add to claude_desktop_config.json:
{
"mcpServers": {
"poe2-build": {
"command": "C:\\Users\\<you>\\AppData\\Roaming\\Python\\Python312\\Scripts\\uv.exe",
"args": ["run", "--directory", "W:\\GitHub\\poe2-build-mcp", "python", "-m", "server.main"]
}
}
}
Data & credit
Game data is GGG's IP, redistributed as derived data within established community-tool norms — with credit to the RePoE fork and poe2db. Build numbers come from the PathOfBuilding-PoE2 community fork. This project is not affiliated with or endorsed by Grinding Gear Games.
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
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.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.