BoundaryXP MCP
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@BoundaryXP MCP Check this task plan for dependency cycles and show the evidence trail."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Questions, counterexamples and contributions are welcome. The 7+1 method has assumptions and limits; the research cases do not establish correctness for every task. The MCP implementation, model interpretations and L2/L3 resources can also be wrong. Deterministic calculation does not guarantee correct semantic interpretation. Help test and improve the approach with questions, counterexamples and reproducible evidence.
Questions and ideas · 7+1 counterexamples · MCP problems · L2/L3 semantic issues · Contribute and submit a PR
Install BXP · Find BXP
Linux x64 / Docker is available separately: npm Linux package · Linux installation · 中文安装 · Dockerfile · Linux release. Version 0.4.0-rc.9-linux.1 uses the linux channel, glibc ≥2.36 and compiled B-335 Core. Windows latest/next remain rc.9. Glama verification is tracked separately; this release does not claim platform approval.
Install from npm → · Codex / Cursor setup → · 中文安装 →
21 tools · 17 resources · Windows x64 · Node.js ≥22.18 · rc.9 prerelease. Default analysis uses your current host LLM; no separate model API key is required.
Public entry | What you can find there |
rc.9 package, Chinese/English descriptions and installation on the same page | |
Published rc.9 metadata and npm installation reference | |
rc.7 listing is merged; rc.9 update PR #6087 is pending | |
Live listing and npm link; rc.8 remains displayed, documentation refresh requested November 16 | |
Live listing with repository and npm links; cached rc.8 snippet omits required data path—use our setup guide | |
Live information page; cached rc.5 instructions, no Glama-hosted deployable release |
Checked 2026-11-16. Search BoundaryXP or boundaryxp-mcp. Use npm for the current package; directory caches can lag. “Listed” means discoverable, not a certification. Full directory status and submission options.
MCPwell and ToolCenter also contain directory entries, but no BXP-specific detail URL was verified. Their category/index pages are listed only in the full directory notes, not as direct BXP installation entries here.
Codex desktop: Plugins → Add → Add MCP server. Follow the stdio setup guide. Add plugin marketplace expects a separate plugin-marketplace source; the npm URL, Registry URL and this product repository are not such a source. BXP is not currently listed in the Codex plugin catalog or Cursor Marketplace.
Plugin contribution: OpenAI Community Plugins PR #57 is ready for maintainer review, with the tested Windows x64 wrapper for compiled rc.9 and a controlled comparison of unrelated Grafana failures. punkpeye PR #19484 has also been updated to rc.9. Both await approval. Test scope and current submission status.
To try the contributor fork through Add plugin marketplace, use the candidate plugin installation guide. Its GitHub-to-Codex installation and installed MCP were verified in isolation; official marketplace approval is still pending.
Related MCP server: Fieldkit
Start with a real case
In ENG20-003, five tasks span several roles. Their combined dependencies form a strict A → B → C → D → A cycle. BoundaryXP brings those dependencies into one graph and preserves the original English sources and calculation records for inspection.

Different cases, different results
Result / case | What to inspect |
Violation · RT044 | Local orders combine into a strict dependency cycle |
Within bounds · RT042 | Adjacent half-open 10–11 and 11–12 bookings release the endpoint |
Incomplete information · L23-050 | “Confirmed by him” has no unique owner binding; U1 remains unknown |
Two cases have original English sources: ENG20-001 · Branches and a join (Within bounds) and ENG20-003 · Four-step dependency cycle (Violation).
All five come from saved 58618 records, retaining task trees, exact sources, verdicts and seven-step reports. Reports and Markdown exports follow the top-right language switch; source evidence keeps its original language. The demo replays B-332 records without connecting to a local service or running new analysis. B335 upgrades do not regenerate those records. Demo and saved reports


Connect to your AI workflow
New in rc.9: import and validate custom L2 industry and L3 enterprise dictionaries through MCP or the workbench. Format and English examples. Existing rc.8 installations can upgrade directly using the rc.9 bootstrap installer once; the old updater's fixed 18-tool probe cannot accept the new 21-tool catalog. Follow the upgrade guide and preserve your existing installation/data paths.
BoundaryXP MCP connects through standard stdio to Cursor, Codex and compatible hosts, with 21 tools, 17 resources and a local browser workbench. The Windows x64 package includes the compiled native Core; no separate 58618 installation or hosted BoundaryXP server is required.
Install from npm: use
boundaryxp-mcp@0.4.0-rc.9, available through thenextprerelease channel. GitHub Releases retain the matching archive and SHA-256.Install and connect: follow the guide to configure the host's stdio MCP server.
Use your current host LLM: the built-in host profile prepares instructions and accepts the host analysis through standard MCP tools. No separate model API key is required; an independent API profile remains optional.
Open the workbench: import complete task text, preview and confirm the task tree, then analyze and inspect evidence and reports.
English installation guide → · 中文安装 · Tools and resources · Updates, rollback and troubleshooting
Current npm release: boundaryxp-mcp 0.4.0-rc.9, B-335 / Windows x64, MIT. Start with the installation guide, 7+1 introduction, 21 tools and directory status. This product repository is public. Core implementation source remains in a separate private repository.
L2 / L3: industry and enterprise semantics
The same term can mean different things across industries and organizations. L2 / L3 provide selected domain context for interpreting tasks alongside the original evidence.
Semantic library | Purpose |
L2 · Industry | Describes domain actions, entities, roles, resources and related terminology to help interpret requirements in their industry context. |
L3 · Enterprise | Adds organization-specific terminology, organizational and role names, controlled aliases and other context to distinguish local objects and expressions. |
There is no active L1 layer. L0 is the BXP protocol and deterministic Core, not a semantic library. L2 / L3 do not replace source text or change the shared 7+1 verification rules. Loading a library, a job title or a reporting relationship does not grant business authority.
Use through MCP
Call
bx_dictionaries_listto inspect the L2 / L3 resources available in the current installation and their identifiers and versions.Call
bx_dictionaries_bindto validate the selected resources and obtain version-and-digest references, keeping library snapshots distinct.Call
bx_dictionary_lookupto look up entries in bound resources. Lookup does not make a semantic ruling or submit analysis.
Ask your host: “List the available BXP L2 and L3 semantic libraries with their identifiers, versions and digests. Wait for my selection before binding; do not submit analysis.”
Install and connect MCP → · Dictionary tool reference · Resource and distribution notices
Available resources are determined by the installed inventory. Manage private enterprise semantic material within its own usage scope. This section documents capabilities and entry points; it does not claim a separate public GitHub library release.
Method and research evidence
BXP: A Method for AI Task Boundary Verification Using a 7+1 Relational Representation
Wei Zhu · English v1.2 / Chinese r23 · First published 2026-11-06
Research cases | Task nodes | Calculation items |
625 | 3,219 | 3,469 |
The paper organizes requirements through 7+1 views and shared relations, uses BXP conventions to establish a common mathematical boundary, and checks whether subtasks and their combination exceed the parent task's boundary. Formal arguments state the conditions for the method. These figures describe research coverage, not accuracy. Case-level agreement, item-level differences and incomplete calculations are reported separately.
English PDF · 中文 PDF · Public research package · Materials guide
DOI: 10.5281/zenodo.24072738. Deposited on Zenodo on 2026-11-09, retaining the first GitHub publication date. GitHub paper release and checksums
The RESEARCH_ONLY_1 supplement contains task sources, annotations, 7+1 reports, operands, results, protocol texts, screenshots and independent research verification scripts. It excludes workbench engineering source, runtime dependencies and production prompt templates. It is maintained separately from the MCP test package.
How 7+1 connects the requirements
I Inputs · O Outputs · A Actions · P Responsibilities and division of work · R Resources · V Version / basis · C Completion / stopping, plus independent B Prohibitions.
Shared relations connect objects, conditions, dependencies and handoffs across task boundaries. Missing information stays missing; a role title does not grant authority.
flowchart LR
A["Complete task text"] --> B["Confirm parent–child task tree"]
B --> C["LLM joint analysis"]
D["Selected L2 / L3 dictionaries"] --> C
C --> E["Shared deterministic Core"]
E --> F["Evidence, findings and unresolved items"]
F --> G["Human / host review"]The B335 workbench makes one semantic request, followed by Core checks, with no separate audit request or automatic retry. LLMs interpret source meaning; Core validates submitted structures, source references, bindings and supported finite operands. Core does not call models, access the network or execute task instructions.
Complete task trees: whole-document UTF-8
.txt/.mdimport, role headings, explicit task and parent IDs, forward references, multiple levels and relationship preview.Inspectable results: task tree, 7+1 radar, model/calculator conclusions, quotations, unresolved items, saved history and exports. Failed attempts remain distinct from saved results.
L2 / L3: versioned industry and enterprise dictionaries; no active L1 layer. Loading resources does not grant business authority.
Chinese and English: interface, reports and Markdown exports preserve original sources and saved verdicts. First-time translation calls a model and is cached by snapshot.
Mathematical input and unknowns: allowed values, uncertain facts and parent/child plan consistency have distinct meanings. Set identities support strings and safe integers without conflating their types or filling in unknown conclusions.
Startup and recovery: historical jobs load with bounded concurrency under existing recovery rules. Readiness follows HTTP identity verification.
Individual sources and serialized joint input are each limited to 1 MiB; HTTP request bodies to 8 MiB. Oversized input is rejected without truncation; there is no 300-code-point ceiling. See the changelog for features and historical validation scope.
Exact quotations do not prove semantic entailment. A completed job does not mean every proof passed. People and host systems retain business decisions and execution authority; business_decision and execution_authorization remain null.
The reviewed baseline is BXP-0001 0.4.0-draft.2; the workbench still uses a recorded draft.1 attachment. Full migration and protocol conformance are not claimed. Implementation notes · Research evidence
Host analysis and first-time translation may send selected source and dictionary content to the configured model provider. Review this data flow before supplying private material.
Explore further
Looking for | Start here |
Industry and enterprise semantics | |
MCP installation and usage | |
Method and citation | |
Protocol and implementation differences | |
Version history and contribution |
Distribution scope: the Git tree contains paper, protocol, cases, reports, screenshots, independent research verification scripts and MCP documentation. It excludes private workbench implementation source and source-build entry points. Compiled MCP packages are published to npm and mirrored in this repository's Releases. Original Core implementation source is excluded from those packages; adapter, UI and required resources remain readable. Full scope
Public status: the owner has made this product repository public. Historical audit notes describe their recorded state; public visibility alone does not confirm that old GitHub references or caches have been cleared. Audit record
The existing MIT license is retained. Paper, case data and third-party materials have their own usage terms; the MCP package has separate license and third-party notices.
Available Tools
21 toolsbx_analysis_submitAIdempotent
默认host模式:prepare取完整规范,当前LLM分析后complete提交JSON,由原生Core计算保存;不需要额外模型API。 / DEFAULT HOST WORKFLOW: use profile_id host. action prepare freezes project/revision and returns instructions_artifact and preparation_sha256. Read ALL instruction pages, analyze with the current LLM, then action complete supplies response (full JSON text), job_id and preparation_sha256. Native Core validates and saves; poll bx_job_get and read result_revision. No extra API or server model call. Legacy API mode requires confirm_model_call:true and an explicit API profile.
Inputs: revision (required), project_id (required), request_id (required), action, +4 more. Returns: data, kind, product, api_version, +2 more
bx_artifact_getARead-onlyIdempotent
按摘要分页读取完整工件,offset单位是Unicode码点。 / Read a digest-bound artifact by Unicode codepoint pages, preserving full source, proposal, verification and report bytes. Reassemble every page and verify SHA256.
Inputs: sha256 (required), limit, offset. Returns: data, kind, product, api_version, +2 more
bx_capabilitiesARead-onlyIdempotent
查看版本、合同及中英帮助;language选zh-CN或en,help_topic选overview/tools/examples/troubleshooting;工具主题可用tool_name查询单个工具。 / Discover the pinned runtime, contracts and bilingual built-in help. Use language zh-CN or en and help_topic overview, tools, examples or troubleshooting; optional tool_name selects one tool in the tools topic. No model call.
Inputs: language, tool_name, help_topic. Returns: data, kind, product, api_version, +2 more
bx_dictionaries_bindARead-onlyIdempotent
绑定显式字典选择,取得固定摘要。 / Validate an explicit L2/L3 selection and return its immutable digest. No task authority.
Inputs: selection (required). Returns: data, kind, product, api_version, +2 more
bx_dictionaries_listARead-onlyIdempotent
列出L2/L3字典、版本、摘要和测试标记。 / Discover B-335 L2/L3 IDs, versions, digests and test-only flags. No active L1.
Inputs: no named parameters. Returns: data, kind, product, api_version, +2 more
bx_dictionary_formatARead-onlyIdempotent
读取自定义L2/L3的格式规范、JSON Schema与完整中英文示例。 / Get the custom L2/L3 JSON Schema, format requirements and a complete Chinese or English example. No files or model required.
Inputs: layer, language. Returns: data, kind, product, api_version, +2 more
bx_dictionary_importAIdempotent
按校验摘要导入自定义字典,同版本不同内容拒绝覆盖;不修改任务或旧字典。 / Import a validated L2/L3 dictionary into this installation data directory. Supply the sha256 from validate. Idempotent by content; conflicting same ID/version is rejected. Never changes tasks or old dictionaries.
Inputs: sha256 (required), document (required). Returns: data, kind, product, api_version, +2 more
bx_dictionary_lookupARead-onlyIdempotent
在已绑定字典中分页查询词条,不作语义裁决。 / Look up lexical entries only in the explicitly digest-bound dictionaries. Not semantic interpretation or authorization.
Inputs: query (required), dictionaries (required), offset, pack_id. Returns: data, kind, product, api_version, +2 more
bx_dictionary_validateARead-onlyIdempotent
校验完整JSON字典,返回字段错误、歧义提示和导入摘要;不保存。 / Validate a complete UTF-8 JSON dictionary document without saving. Returns field errors, ambiguity warnings and content sha256 for import. No model call.
Inputs: document (required). Returns: data, kind, product, api_version, +2 more
bx_job_cancelBIdempotent
请求取消指定作业,保留历史,不保证远端立即终止。 / Request cancellation; retain history and completed results. Remote termination is not guaranteed.
Inputs: job_id (required), project_id (required). Returns: data, kind, product, api_version, +2 more
bx_job_getBRead-onlyIdempotent
查询作业状态、失败、调用数及结果版本。 / Read one owned job status, failures, observed request count and exact saved result reference. No model call.
Inputs: job_id (required), project_id (required). Returns: data, kind, product, api_version, +2 more
bx_profiles_listARead-onlyIdempotent
列出模型配置及就绪状态,不返回密钥。 / List locally configured model profiles and readiness. No key values or model call.
Inputs: no named parameters. Returns: data, kind, product, api_version, +2 more
bx_project_createBIdempotent
保存新项目,保留原文并返回项目ID与版本。 / Save a full parent/child task as a new isolated B-335 project. Idempotency key binds exact input. Does not call a model.
Inputs: input (required), request_id (required). Returns: data, kind, product, api_version, +2 more
bx_project_getARead-onlyIdempotent
读取指定项目的精确版本及完整内容引用。 / Read exact project revision metadata, source digests and artifact references. Revision required; no latest fallback.
Inputs: revision (required), project_id (required). Returns: data, kind, product, api_version, +2 more
bx_projects_listARead-onlyIdempotent
分页列出已保存项目。 / List isolated projects without source text. Paginated; no model call.
Inputs: limit, offset. Returns: data, kind, product, api_version, +2 more
bx_project_updateBIdempotent
按精确版本编辑任务或设置,保留历史,不调用模型。 / Apply one validated task edit to an exact revision, preserving old snapshots. Does not call a model; results become stale until a new analysis.
Inputs: command (required), revision (required), project_id (required), request_id (required). Returns: data, kind, product, api_version, +2 more
bx_result_getARead-onlyIdempotent
读取指定结果版本的LLM、计算器、差异和报告。 / Read the joint LLM/Core summary and immutable result/report references for an exact revision. No rerun or translation.
Inputs: revision (required), project_id (required). Returns: data, kind, product, api_version, +2 more
bx_result_recheckBIdempotent
复算保存的原候选,生成新版本,零次新模型调用。 / Recompute saved submitted calculations using the same B-335 Core. Preserve the original LLM proposal and old results; create a new revision with zero model requests.
Inputs: revision (required), project_id (required), request_id (required). Returns: data, kind, product, api_version, +2 more
bx_task_importBRead-onlyIdempotent
预览完整Markdown父子任务,支持多个DEV和明确父任务编号,不自动保存。 / Preview complete Markdown parent-child input. Supports ## PM → DEV(1/14), repeated DEV roles and explicit Task ID/Parent Task ID. Missing parents require confirmation; numbering never implies order. Does not create a project or call a model.
Inputs: label (required), source (required), profile_id (required), dictionaries (required), +1 more. Returns: data, kind, product, api_version, +2 more
bx_task_validateARead-onlyIdempotent
校验完整结构化输入、父子关系及容量,不保存。 / Validate full source texts, explicit parent links, resource binding and joint byte limits without persisting or calling a model.
Inputs: input (required). Returns: data, kind, product, api_version, +2 more
bx_workbench_openARead-onlyIdempotent
返回独立工作台当前地址,由宿主打开,不启动分析。 / Return the independent B-335 workbench URL. Optional project link; browser opening is controlled by the host. Does not start analysis.
Inputs: project_id. Returns: data, kind, product, api_version, +2 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 21 tools
Most tools map to distinct resource+action pairs (project create/get/update, job get/cancel, result get/recheck, artifact get), and the dictionary family of six tools is differentiated by descriptions (discover / bind / lookup / format / validate / import). The riskiest overlaps are the preview-only vs persisting tools (bx_task_import, bx_task_validate, bx_project_create) and bx_project_update vs bx_task_validate, which an agent could misselect without reading the long descriptions carefully.
Names follow a stable bx_<resource>_<verb> snake_case pattern throughout (bx_project_create, bx_job_get, bx_result_recheck, bx_dictionary_import), which is readable and predictable. Minor deviation: singular/plural resource mixing within the same family (bx_projects_list vs bx_project_get; bx_dictionaries_list / bx_dictionaries_bind vs bx_dictionary_lookup / bx_dictionary_format).
21 tools is on the heavy side for the surface; each phase (discovery, projects, dictionaries, analysis, jobs, results, artifacts) has justification, but the dictionary sub-domain alone consumes six tools. The set is coherent rather than bloated, yet sits in the borderline 16-25 band.
Core lifecycle is largely covered: project create/get/update/list, analysis prepare/complete, job get/cancel, result get/recheck, artifact paging. However there is no project delete/archive, no job or result listing/history enumeration tool, and task-level editing only exists indirectly through bx_project_update, leaving dead ends for cleanup and discovery of prior runs.
Maintenance
Related MCP Connectors
Persistent memory, hybrid search and a goal graph for AI agents, over stdio or remote HTTP.
Design, build and run teams of AI agents with checked handoffs between them.
AI-native programming language and toolchain for accountable, verifiable agent-driven development.
Pre-execution governance for AI agents. Deterministic PASS/FAIL/REVIEW verdicts, replayable proof.
Related MCP Servers
- AlicenseNot gradedqualityFmaintenanceEnables AI agents to perform auditable A* reasoning, import world graphs, incorporate physical carrier constraints, and maintain a learning knowledge base through a zero-dependency local MCP stdio service.MIT
- AlicenseBqualityBmaintenanceEnables a small local AI model to discover, describe, run, undo, and verify deterministic tools over JSON-RPC on stdio, with previews and approval gates for changes.10AGPL 3.0
- AlicenseNot gradedqualityCmaintenanceEnables AI coding agents to inspect and reverse-engineer large binaries through bounded stdio JSON-RPC 2.0 tools for graph queries, SSA facts, path solving, runtime grounding, and mod scaffolding.1AGPL 3.0
- AlicenseNot gradedqualityCmaintenanceBridges a planning and review client to DeepSeek Harness Desktop over local stdio MCP, letting users submit staged tasks with explicit goals, scope, and acceptance criteria for execution inside native DSH sessions, models, and approval boundaries. Exposes task submission, waiting, evidence retrieval, review, continuation, cancellation, and recovery tools so results are judged against traceable evidence from the current run.MIT