Blawx Server

A minimal run-local MCP server that calls the Blawx API using a Blawx API key.

Prereqs

• Python 3.10+

Install

From this repo root:

python -m pip install -e .

Example: Windows + Claude Desktop

This is one way to set up the MCP server locally for Claude Desktop on Windows.

1. Clone the repository locally.

git clone https://github.com/Lexpedite/blawx_mcp.git
cd blawx_mcp

2. Install the package.

python -m pip install .

3. Generate an API key from the Blawx profile page. In the left navigation bar, click "Profile", then use the "Add API Key" button. Copy the key when it is shown.

4. Add the MCP server to your Claude Desktop configuration file, typically at %APPDATA%\Claude\claude_desktop_config.json:

{
	"mcpServers": {
		"blawx-mcp": {
			"command": "python",
			"args": ["-m", "blawx_mcp", "--stdio"],
			"env": {
				"BLAWX_API_KEY": "your API Key Here"
			}
		}
	}
}

The MCP server no longer needs team_slug or project_id at startup. Agents discover teams at runtime with blawx_teams_list, choose a team slug, then discover projects with blawx_projects_list.

You will need to completely quit Claude Desktop (using the Claude icon in the system tray and selecting "Quit", then restarting) and restart for the changes to your configuration to take effect.

Configuration

Set required configuration in your environment:

export BLAWX_API_KEY="your_key_here"

Team and project selection happen at tool-call time instead of startup time. Use blawx_teams_list after connecting the server. If it returns exactly one team, use that team's slug as team_slug; if it returns multiple teams and the user has not already identified one, ask the user which team to use. Then call blawx_projects_list with team_slug and pass both team_slug and the returned project_id to every project-scoped tool.

You can create a Blawx API Key if you have a Pro subscription to Blawx. Click on "Profile" in the left navigation bar, and find the "Add API Key" button. When you click the button your API key will be displayed only once at the top of the screen. Copy and paste it into your environment settings.

Run

Run the MCP server from this folder (no install required).

For SSE/HTTP transport:

./.venv/bin/python -m blawx_mcp

Defaults:

• Binds to 127.0.0.1:8765
• SSE endpoint at http://127.0.0.1:8765/sse

Optional server bind overrides:

export BLAWX_MCP_HOST="127.0.0.1"
export BLAWX_MCP_PORT="8765"

For stdio transport, which is useful for local clients such as Claude Desktop:

./.venv/bin/python -m blawx_mcp --stdio

Connect to Your Coding Agent

Coding agents differ in how they configure MCP servers. This is a typical tool definition in your mcp.json for VS Code.

{
	"servers": {
		"my-blawx-sse-server": {
			"url": "http://127.0.0.1:8765/sse",
			"type": "http"
		}
	},
	"inputs": []
}

Tools

These tools give your coding agent the following capabilities:

All project-scoped tools require both team_slug and project_id. Agents should always call blawx_teams_list first. If there is only one team, use that team's slug; if there are multiple teams and the user has not already identified one, ask the user which team to use. Then call blawx_projects_list with that team_slug and pass both team_slug and the returned project_id to every later ontology, legal-doc, question, fact-scenario, ask/answer, and encoding tool.

1. Discover teams and choose a team_slug.
2. Discover projects under that team and choose a project_id.
3. Discover what the chosen project exposes (questions, fact scenarios, ontology).
4. Create, update, and delete LegalDocs, LegalDocParts, and EncodingParts.
5. Ask a question (using either a stored fact scenario or a custom facts payload).
6. Browse answers and drill into explanations (model/attributes/explanation text).

Here's a brief run-down of the available tools.

Health check

• blawx_health: verifies the Blawx app is reachable and returns status/body.

Discover Project Content

Agents should start by listing teams, then listing projects under the selected team.

• blawx_teams_list: lists the teams available to the configured API key.
• blawx_projects_list: lists the available projects for a team_slug.
• blawx_project_detail: retrieves metadata for a specific project id.

Once a team and project are chosen, every project-scoped tool requires explicit team_slug and project_id arguments. Agents will then usually list the available questions, fact scenarios, and vocabulary.

• blawx_questions_list: lists shared questions available in the project.
• blawx_question_detail: retrieves a specific question's details (useful when deciding which question id to ask).
• blawx_fact_scenarios_list: lists stored fact scenarios (prebuilt sets of facts you can re-use).
• blawx_fact_scenario_detail: shows the facts contained in a specific fact scenario.
• blawx_ontology_list: lists ontology categories/relationships (the project's vocabulary).
• blawx_ontology_category_detail: details for a specific category.
• blawx_ontology_relationship_detail: details for a specific relationship (including arity/parameters).

Additional read-write tools are also available for project editing (questions, fact scenarios, ontology categories/relationships/parameters).

Examples:

{
	"team_slug": "my-team",
	"project_id": 60
}

{
	"team_slug": "my-team",
	"project_id": 60,
	"question_id": 123
}

For write operations:

• blawx_encodingpart_update, blawx_question_create, blawx_question_update, blawx_fact_scenario_create, and blawx_fact_scenario_update all use the same payload shape:

{
	"payload": {
		"blawx_json": {
			"...": "..."
		}
	}
}

For question saves specifically, the encoding is expected to include a single outer question block.

• Ontology write tools accept plain JSON objects under payload, not blawx_json workspaces. Current API validation requires these minimum shapes:

{
	"payload": {
		"name": "Contract",
		"slug": "contract",
		"short_description": "",
		"nlg_prefix": "",
		"nlg_postfix": "is a contract"
	}
}

{
	"payload": {
		"name": "Estimated Expenditure",
		"slug": "estimated_expenditure",
		"short_description": "",
		"nlg_prefix": ""
	}
}

{
	"relationship_id": 458,
	"payload": {
		"order": 1,
		"type_id": 466,
		"nlg_postfix": ""
	}
}

Notes:

• Category create/update currently requires name and slug; nlg_postfix must be 50 characters or fewer.
• Relationship create/update currently requires name and slug.
• Relationship-parameter create/update currently requires order and type_id.
• Use blawx_ontology_categories_list or blawx_ontology_category_detail to discover valid category ids for type_id.

Patch-style tools are intentionally not exposed in this MCP server to reduce tool-selection ambiguity.

Ask Questions

• blawx_question_ask_with_fact_scenario: asks a question using a stored fact scenario.
• blawx_question_ask_with_facts: asks a question using an explicit facts payload generated by your agent based on your instructions.

Both ask tools require team_slug and project_id in addition to question_id and their respective payload arguments.

Both ask tools currently require shared questions. If you call either blawx_question_ask_with_fact_scenario or blawx_question_ask_with_facts with a non-shared question id, the Blawx app may return Question not available via API. Use a question from blawx_questions_list, or set shared: true on the question first.

NB: It's not clear how good agents will be at generating representations of complicated fact scenarios in complicated vocabularies. It can be helpful to review how your agent formulated your fact scenario if you get unexpected results, and to give it hints on how to do better.

When you pose a question, the answer is saved on the Blawx server for approximately 30 minutes, and your agent can review it over that period of time. Once the data expires, your agent will need to pose the question again to analyse the responses further. Based on the instructions provided by the MCP server, it should know to do that when and if required.

Review Answers

Blawx's answers can be quite large, and agents have a limited context window, so the process of reviewing answers is broken into multiple steps.

1. Get the list of answers to the question.
2. Get the list of explanations for a specific answer.
3. Look at the parts of a specific explanation.

• blawx_list_answers: gives the list of answers available, and the bindings in those answers.

• blawx_cached_response_meta: metadata (ttl, created, answer_count) for a cached response.

• blawx_list_explanations: gives the list of explanations available for an answer

There are four tools to retrieve specific parts of an explanation. These tools all allow the agent to select the entire part, or if it is too long, to select only certain lines at a time.

• blawx_get_model_part: this returns the answer set
• blawx_get_attributes_part: this returns the constraints applied to variables in the model and explanations
• blawx_get_explanation_part: this is the tree-structured, human-readable explanation for the answer
• blawx_get_constraint_satisfaction_part: this is the portion of the explanation that shows how global constraints were satisfied. This is often both verbose and unhelpful, so it is separated out. You may need to ask your agent to seek it specifically if you know your encoding uses constraints and you need to know how they are satisfied.

Legal Docs + Encoding

The API supports read-write legal documents, legal document parts, and encoding parts, and this MCP server now exposes CRUD tools for all three layers.

• blawx_legaldocs_list, blawx_legaldoc_create, blawx_legaldoc_detail, blawx_legaldoc_update, blawx_legaldoc_delete
• blawx_legaldocparts_list, blawx_legaldocpart_create, blawx_legaldocpart_detail, blawx_legaldocpart_update, blawx_legaldocpart_delete
• blawx_encoding_guide, blawx_encodingpart_get, blawx_encodingpart_update, blawx_encodingpart_delete

All of these tools require team_slug and project_id.

Default LegalDocPart granularity rule: preserve the legislative hierarchy. Each heading gets its own part, each section gets its own part for section-level text, a section with no subordinate units is one part, and each subsection, paragraph, or similar subordinate unit gets its own part when it has distinct text. Do not merge separately addressable legislative units into one part just because their logic is related.

For the full LegalDocPart structure guidance, including parent_id, include_parent, include_sibling, and worked examples, read blawx_encoding_guide topic legaldocs.

To read legislation text, use this sequence:

1. blawx_teams_list to choose a team_slug.
2. blawx_projects_list to choose a project_id.
3. blawx_legaldocs_list (or blawx_legaldoc_detail) to identify a legal_doc_id.
4. blawx_legaldocparts_list to list part ids/titles/order for that document.
5. blawx_legaldocpart_detail for each relevant legal_doc_part_id to retrieve the actual part text/content.

blawx_legaldocparts_list is primarily navigational metadata; blawx_legaldocpart_detail is the tool that returns the text for a specific part.

To create or update document structure:

1. blawx_encoding_guide with topic legaldocs.
2. blawx_legaldoc_create or blawx_legaldoc_update with payload:

{
	"team_slug": "my-team",
	"project_id": 60,
	"payload": {
		"name": "Financial Administration Act",
		"slug": "FAA",
		"tag_ids": []
	}
}

3. blawx_legaldocpart_create with payload such as:

{
	"team_slug": "my-team",
	"project_id": 60,
	"legal_doc_id": 78,
	"payload": {
		"element_type": "Section",
		"index_text": "41",
		"text_content": "No contract shall be entered into...",
		"substantive": true
	}
}

For part updates, omit parent_id. The current Blawx API rejects re-parenting via PUT.

Use this sequence first:

1. Call blawx_encoding_guide (topic quickstart, then encoding-process, then encodingpart, then blawx-json). Additional available topics include:
   • valid-blawx-json: concrete valid payload patterns for blawx_json
   • blawx-blocks: quick reference of available block types and required components
   • encoding-process: step-by-step workflow for creating and updating encoding parts
   • legaldocs: LegalDoc and LegalDocPart structure, write fields, and project-selection workflow
2. Call blawx_encodingpart_get to inspect existing encoding.
3. Call blawx_encodingpart_update with payload shape:

{
	"payload": {
		"blawx_json": {
			"...": "..."
		}
	}
}

blawx_encodingpart_update accepts only Blawx JSON blocks via payload.blawx_json. Do not send content, scasp_encoding, or stringified JSON.

NB: The other three parts should be read alongside the attributes, or relevant information may be missing. This instruction is provided to the agent, but if it isn't followed your agent may draw incorrect conclusions. It may be wise to instruct your agent to check the attributes in addition to the other parts of an explanation.

Typical Workflow

1. Start the server with BLAWX_API_KEY.
2. Call blawx_teams_list and choose a team_slug; if only one team is returned, use it.
3. Call blawx_projects_list with team_slug and choose a project_id.
4. Pass that team_slug and project_id to ontology, fact-scenario, question, ask/answer, legal-doc, legal-doc-part, and encoding-part tools.

Development

The Blawx server used can be overridden for local development

• BLAWX_BASE_URL (default: https://app.blawx.dev)

Embedding as a Library (multi-tenant / hosted use)

blawx_mcp can be imported as a library in a hosted multi-tenant server. Use settings_context() to inject per-request Settings so concurrent requests from different users never collide on configuration or cached state.

from blawx_mcp import Settings, settings_context
from blawx_mcp.config import _settings_override

settings = Settings(
    base_url="https://app.blawx.dev",
    api_key="my-key",
)
token = settings_context(settings)
try:
    # ... invoke MCP tools in this async context ...
    pass
finally:
    _settings_override.reset(token)

settings_context() uses Python's contextvars.ContextVar under the hood, which is async-safe: each asyncio task inherits its own copy of the context so concurrent requests cannot collide even when running in the same event loop. Team selection is provided by MCP tool arguments, so hosted consumers should pass team_slug through the team-scoped tool calls rather than injecting it into settings.