Gridline — AEC Architecture Agent Team (12 Specialists, 30 Skills)

From brief to blueprint to handover.

30 portable AI Agent Skills + 12 specialist agents for architecture, BIM, CAD and multidisciplinary AEC workflows.

Copy the install prompt · Current status · Skills · Drawings · Tools · Research

https://github.com/user-attachments/assets/a397a1b3-0f50-4d00-948a-62c9537144d8

---

Gridline gives capable AI coding/agent tools a structured way to help with the full building-project lifecycle: programming and feasibility, site analysis, concept design, plans/sections/elevations, BIM/CAD workflows, structural/MEP/civil coordination, technical documentation, performance, quantities/cost, construction administration, handover and post-occupancy.

It is not one giant “act like an architect” prompt. Each specialist is discovered only when relevant, keeping the always-loaded context small and the workflow auditable.

What Gridline actually ships — October 2026

Gridline combines 30 portable AEC skills and 12 specialist agents with runnable drafting and review utilities. The descriptions below distinguish implemented software, partial review coverage, and planned capabilities. Architectural validation is not professional certification.

Capability	Current evidence and scope
Agent installation and skill routing	Cross-provider packages, portable skill directories, a one-copy installation prompt, provider activation fixtures, and reversible/ownership-aware global installers (v0.5.0–0.5.1)
Conceptual drawings	Dependency-free JSON-to-SVG/DXF engine for six supported primitives; output remains conceptual, not for construction (drawing details)
Specification-derived CAD plan sheets	Structured building JSON → standard-layered DXF plans, paper-space layouts/title blocks, required vector PDFs, room/door/window schedules and a sheet register, with ISO 13567 / NCS profiles and office overrides (PR #6)
Local drawing checks	aec_cad.py check independently measures supported wall/opening/room primitives and PDF straight-vector plot mapping, plus selected drafting/title/tag/schedule invariants. Receipts bind input/output/validator hashes and coverage. This is supported local 2D verification, not complete geometry, PDF visual fidelity or native AutoCAD acceptance. (verification scope)
Read-only revision-impact review	Compare two building specs to list changed openings, walls, rooms, levels, grids and document settings, affected indexed sheets and impacted schedules. Coverage is partial: declared host/level/receiving-room links, missing model references and grid/document-setting changes are tracked; report output is atomic and rejects input aliases, but room-wall adjacency and unrelated CAD/PDF drawings are not independently inspected (PR #9, issue #10)
Local read-only MCP server	Optional Gridline-owned stdio MCP review server exposes revision comparison and capability discovery. It does not connect to or edit AutoCAD, Revit or other native applications (PR #9)
Professional host integration	Revit, AutoCAD, Rhino/Grasshopper, IFC/Bonsai and related workflow guidance exists in skills. Native MCP tool connections, host acceptance testing and controlled model writes are not yet shipped (issue #12)
Local Blender 3D proof (experimental)	scripts/blender_native.py can generate a new native .blend + .glb from supported building-spec walls/levels/doors/windows and independently reopen/measure outputs in a locally installed Blender version. Does not create detailed roofs/interiors or reconstruct architecture from a reference photo; this is a CLI bridge, not a native Blender MCP (scope and usage)
Optional ARCHIMESH components (0.9.0)	Typed component catalog, explicit pinned-source execution, disposable Blender 5.2.3 native/GLB readback, signed local proposals and opt-in CC0 texture cache. The local ledger now also rejects a head rewind to an older signed content hash without an approved transition. Source-bound IFC4 and real texture readback are tested; professional/client and live-host gates remain explicit (usage, evidence and limits).

Verified limitations and next engineering gates

The independent Gridline/Codex technical audit reproduced a critical defect in the old checker: deleting one opening LINE while preserving tags/schedules still returned PASS. PR #21 adds independent readback and adversarial tests for supported primitives. Full geometric completeness, raster visibility and native acceptance remain separate gates; the checker is not construction-ready certification.

Other gates remain open: cross-sheet comparison and 40-sheet reference fixtures (#10), IFC schema/IDS validation (#11), real CAD/BIM adapters (#12), editable 2D→3D/BIM (#13), scanned/hand-marked plans (#14), human-approved revision state/transactions (#15), and native AutoCAD open/AUDIT/edit/plot/round-trip evidence (#16).

These are implementation issues, not completed product claims. The community progress ledger records delivered pieces, proof, remaining acceptance criteria, and why issues stay open. Contributions, professional CAD sample sets and critical reviews are welcome.

Install with your AI agent — copy once

No terminal knowledge required. Open Codex, Claude Code, Gemini CLI, Cursor, OpenCode, Copilot, or another filesystem-capable agent and paste this:

Install Gridline for my current user.
Official source: https://github.com/MakeItEzzz555/Gridline

First inspect the official README, INSTALL_PROMPT.md, skills/manifest.json and installer/provider docs. Detect which compatible agent runtimes are already installed.

Prefer the standard Agent Skills installer when supported:
npx skills@latest add MakeItEzzz555/Gridline --all -g

Otherwise use this repository's documented native Codex/Claude/Gemini method or reversible global installer. Preserve existing skills/configuration and back up conflicts. Do not use sudo/root, modify my architecture projects, read credentials, delete unrelated skills, or install third-party MCP servers/plugins.

Install only Gridline for the detected hosts. Verify the installed stable-skill count against skills/manifest.json and use documented provider list/validate commands when available. Report PASS/FAIL per host, installed and backup paths, update/uninstall instructions, and whether I need to reload or start a fresh session.

Stop rather than guessing if a command, path, host or permission model is unclear. Do not claim installation succeeded until the filesystem/provider confirms it.

Full install prompt, alternatives and safety notes.

One command

For Agent Skills hosts supported by the community skills installer:

npx skills@latest add MakeItEzzz555/Gridline --all -g

To opt out of that CLI's documented anonymous telemetry:

DISABLE_TELEMETRY=1 npx skills@latest add MakeItEzzz555/Gridline --all -g

Native provider installs

Codex

codex plugin marketplace add MakeItEzzz555/Gridline
codex plugin add gridline@gridline

Claude Code

claude plugin marketplace add MakeItEzzz555/Gridline
claude plugin install gridline@gridline

Gemini CLI

gemini extensions install https://github.com/MakeItEzzz555/Gridline

Or clone once and let the repository detect installed runtimes:

git clone https://github.com/MakeItEzzz555/Gridline.git
cd Gridline
bash scripts/install-global.sh auto

Windows users can run powershell -ExecutionPolicy Bypass -File .\scripts\install-global.ps1 auto.

Measured skill activation

Record actual Codex, Claude Code or Gemini CLI activations and use the dependency-free behavioral evaluation scorer to measure per-skill recall, negative controls and regressions against a baseline. CI tests the scorer without making model calls or claiming measured routing accuracy.

The 30 superpowers

Brief, research and site

Skill	Use it for
architecture-orchestrator	End-to-end routing and multidisciplinary synthesis
architecture-programming	Briefs, rooms, areas, adjacencies and requirements
site-analysis-feasibility	Survey/context/climate constraints, opportunities and capacity
site-regulation-research	Planning, zoning and building-control source research
precedent-research	Evidence-backed architectural case studies and transferable lessons

Design and drawing

Skill	Use it for
concept-schematic-design	Parti, massing, blocking, plans, sections, elevations and options
concept-design-review	Critique and compare existing schemes
interior-fitout-design	Fit-out, furniture/equipment, RCP, finishes and joinery briefs
architectural-drawing-generation	Site/floor/roof/RCP plans, elevations, sections, details, SVG/DXF
visualization-rendering	Cameras, lighting, materials, render/animation workflows
presentation-review	Boards, narratives, jury/client presentation quality

CAD, BIM and interoperability

Skill	Use it for
revit-workflow	Revit, pyRevit, Dynamo and Revit API
rhino-grasshopper-workflow	Rhino, Grasshopper, RhinoCommon and parametric geometry
autocad-dwg-workflow	AutoCAD, DWG, xrefs, layers, blocks, plotting, AutoLISP/.NET
openbim-ifc-coordination	IFC, BCF, IDS, bSDD, IfcOpenShell/Bonsai and federation
cad-bim-automation	Generic DXF/SVG/Blender/FreeCAD and cross-platform automation

Engineering and performance coordination

Skill	Use it for
civil-site-design	Survey, grading, drainage, utilities, roads/parking and Civil 3D
structural-coordination	Grids, spans, cores, transfers, openings and engineer interfaces
mep-coordination	Plant, shafts, risers, ceilings, services, penetrations and clashes
life-safety-accessibility	Sourced egress, fire/life-safety and accessible-route coordination
building-envelope-performance	Water/air/vapor/thermal layers, facades, roofs and interfaces
sustainability-daylight-energy	Passive design, daylight, energy, carbon, water and resilience

Technical delivery, cost and operation

Skill	Use it for
construction-documentation	Drawing sets, schedules, details, revisions and issue packages
drawing-qa	Cross-sheet drawing QA
materials-specifications	Materials, finishes, performance requirements and specifications
area-quantity-audit	Areas, counts, schedules and takeoff reconciliation
cost-planning-value-engineering	Early cost plans, option deltas and VE registers
construction-administration	RFIs, submittals, site observations, changes and punch lists
handover-fm-postoccupancy	O&M, asset data, closeout, FM and POE
architecture-red-team	Independent final challenge before consequential issue/submission

Full lifecycle map.

Revision impact and cross-sheet QA (read-only pilot)

When an opening, wall, room, level, grid, project unit, drafting standard or drawing revision changes, Gridline identifies affected indexed sheets and schedules for review. It follows declared wall→opening and level→wall/room dependencies in both revisions, including detail-only references. Grid edits and document-setting changes conservatively flag all indexed sheets, since no per-sheet grid/style dependency index exists; unit changes also flag dimensional schedules. The additive changed_document_fields report field identifies unit, standard, schema-version and other top-level document-field changes (including site/datum/material extension fields). Unknown field dependencies trigger conservative sheet review, not a claim of complete schedule coverage. This is read-only source-spec impact analysis; it does not inspect unrelated PDFs, prove complete room-wall adjacency, or certify 40 unknown external drawings.

python3 skills/architectural-drawing-generation/scripts/revision_impact.py before.building.json after.building.json --out impact.json

There is also an opt-in local, read-only Gridline MCP server exposing this comparison to MCP-compatible agents (requires the official Python MCP SDK v2 and an explicitly permitted project directory). It is not a connection to AutoCAD or Revit. On POSIX systems with descriptor-relative no-follow opens, project JSON reads pin the authorized directory and opened regular-file descriptors, rejecting symlink swaps and reads beyond 5 MB. On Windows and platforms without those primitives, the compatibility fallback retains path validation and bounded reads but cannot guarantee protection against concurrent pathname replacement; use a trusted, non-writable project directory there.

Next phases: IFC/IDS validation, official Autodesk MCP, Rhino, Blender, Archicad, editable 3D and scanned drawing research.

Coordinated CAD drawings

When a trusted native CAD/BIM host is available, the skills provide guidance for working with the native editable model; Gridline does not currently provide an installed native-host connection. Without one, the drawing skill can build a specification-derived plan-sheet set: standard-layered DXF files (ISO 13567, US NCS or an office profile), required vector PDFs, room/door/window schedules and a sheet register. The current checker inspects a defined subset of properties and schedule consistency; it does not certify geometry completeness, PDF fidelity, cross-sheet views or AutoCAD editability. Native AutoCAD open/AUDIT/edit/plot/round-trip testing is an outstanding acceptance gate (#16).

pip install ezdxf==1.4.5 matplotlib==3.11.2 pypdf==6.21.2
D=skills/architectural-drawing-generation
python3 $D/scripts/aec_cad.py build $D/references/example-house.building.json /tmp/example-house
python3 $D/scripts/aec_cad.py check $D/references/example-house.building.json /tmp/example-house

Example sheet A-101: ground floor plan at 1:100 on A3, built and checked from the example spec

It draws plans today (walls, doors, windows, rooms, grids, dimensions, sheets); sections, elevations and details are next. For quick diagrams there is also a small dependency-free engine that turns an explicit JSON spec into SVG or DXF. It draws six primitives (lines, rectangles, circles, polylines, text and dimensions), enough for conceptual plan and section diagrams. It does not model walls, doors, hatching, title blocks or scale bars:

python3 skills/architectural-drawing-generation/scripts/aec_draw.py skills/architectural-drawing-generation/references/concept-plan.drawing.json /tmp/plan.svg
python3 skills/architectural-drawing-generation/scripts/aec_draw.py skills/architectural-drawing-generation/references/concept-plan.drawing.json /tmp/plan.dxf

The JSON format is documented by aec-drawing.schema.json. The fallback never invents dimensions and stamps conceptual output CONCEPT / NOT FOR CONSTRUCTION by default. Drawing-output details.

A bounded AutoCAD Mac inspection adapter prepares a disposable DXF copy and a manually activated read-only AutoLISP inspector. Installed status, operator-captured metadata and native AUDIT/plot/DWG roundtrip remain distinct evidence gates; this is not an MCP connection.

CAD / BIM / AEC tools

The pack knows how to route work for Revit, pyRevit, Dynamo, Rhino, Grasshopper, AutoCAD, Civil 3D, IFC/IfcOpenShell/Bonsai, DXF/SVG, Blender and FreeCAD. Public MCP projects also exist for Revit, Rhino/Grasshopper, AutoCAD, Archicad, Vectorworks and SketchUp.

Those integrations are optional and not installed automatically. A model-writing bridge can change professional project files; review its source, permissions, network surface and write commands before connecting it.

Tool-integration policy.

A dependency-free source-detail intake gate records unknown dimensions and approvals before exact 3D work. The incomplete villa example blocks plot boundary, floor-to-floor heights and roof; concept remains provisional. This is an intake check, not model completeness or native-write permission.

A real project context

Copy templates/ARCHITECTURE_PROJECT.md into a project root and fill only what is known. Additional reusable templates cover:

• design brief and room/program inputs;
• site analysis;
• drawing briefs and sheet registers;
• BIM coordination;
• RFIs and site observations;
• punch/defects tracking;
• handover/closeout.

Then ask naturally:

Develop three schematic options from this brief and site evidence, draw the preferred plan and section, reconcile its area schedule, and identify the structural/MEP/life-safety questions that must be resolved next.

The orchestrator should load only the relevant specialists.

How the system works

request
  │
  ▼
architecture-orchestrator
  ├── establishes stage / jurisdiction / units / source truth
  ├── routes only relevant skills
  ├── prefers native CAD/BIM tools when safely available
  ├── uses bounded specialist agents for independent review
  └── verifies and red-teams consequential outputs

The design principle is progressive disclosure: short skill names/descriptions support discovery; full workflows load on activation; scripts, templates and reference docs are opened only when needed.

Professional and safety boundary

Architecture, structural engineering, civil engineering, MEP, fire engineering, accessibility, cost consultancy and construction administration overlap—but they are not interchangeable professional responsibilities.

This project can research, calculate, draft, automate, coordinate, generate conceptual drawings and review evidence. It does not by itself provide licensure, statutory approval, signed/stamped engineering, permit approval or issue-for-construction certification.

The core rules are deliberately conservative:

• never invent survey coordinates, dimensions, regulations, loads/capacities, product certifications or current costs;
• cite jurisdiction, authority, source and edition for regulatory claims;
• preserve source models and issued files;
• use read-only inventory and dry runs before bulk writes;
• make units, origin/datum, IDs and revisions explicit;
• distinguish observed fact from assumption and design option;
• require competent human review for regulated and construction-critical conclusions.

Research-driven, not prompt folklore

The capability map is grounded in professional lifecycle frameworks, buildingSMART/openBIM standards, current vendor documentation and public AEC agent/tool projects. The research specifically informed:

• lifecycle coverage from briefing through post-occupancy;
• version-aware Revit/Rhino/AutoCAD workflows;
• IFC + BCF + IDS information coordination;
• Civil 3D/site workflow concepts;
• guarded live-tool/MCP write patterns;
• progressive skill disclosure instead of a huge permanent prompt.

Research basis.

Maintainer validation

python3 scripts/sync-provider-skills.py --check
python3 tests/check_repo.py
python3 -m unittest discover -s tests -v
bash scripts/validate-native.sh

The repository tests structural skill coverage and provider synchronization. Native provider validation runs only for installed CLIs. Behavioral model activation still requires provider/model evals; installation alone is not proof that every model will select every skill perfectly.

Project status

This is an open-source professional-assistance toolkit, not a substitute for appointed professionals. Coverage is broad, but every real office has its own BIM standards, detail libraries, codes, contracts and QA procedures. Those should be layered into project/organization-specific skills rather than hard-coded as universal rules here.

Research · Lifecycle · Drawings · Providers · Roadmap · Security · Contributing

---

Gridline — AEC agent skills, specification-derived drawings and auditable reviews. Native quality gates are still in development.

Community project. Not affiliated with Autodesk, McNeel, Graphisoft, Nemetschek, buildingSMART, OpenAI, Anthropic or Google.

---

TalentDelta provenance record (added at packaging)

• Packaged for: TalentDelta catalogue (draft; gated by review)
• Upstream: MakeItEzzz555/Gridline
• Pinned commit: 623f57b4a63ee449c27d975524f5e7a53e97d9f5 (2026-11-01)
• Licence: MIT AND GPL-2.0-or-later - the repository LICENSE (MIT) ships unchanged; the GPL-2.0-or-later component (skills/architectural-drawing-generation/scripts/components_worker.py, tests/components_native_acceptance.py) ships with its complete source, the full GPL text (references/GPL-2.0.txt) and the upstream BLENDER-LICENSING.md boundary note, so both licences travel with the archive.
• Upstream author credit: Nicolas Savvides (MakeItEzzz555)
• Runtime: Claude Code, Codex CLI or Gemini CLI (portable Agent Skills folders plus plugin manifests); Python 3 for the drawing, DXF and QA utilities. Optional Blender 5.x for the GPL worker.
• Changes made at packaging: this section was appended to README.md. Upstream paths deliberately NOT shipped in this archive: .github/workflows (upstream GitHub Actions workflows (repository CI, not needed to run the agent)). Edited at packaging, so NOT byte-identical to the pinned commit: README.md (the provenance, licence and safety record appended above). Every other file is byte-identical to the pinned commit.
• Safety read of the packaged archive (packaging time, 2026-11-01): scanned every text file: no API keys, tokens, private keys or .env files; the only credential-shaped strings are test fixtures and documentation placeholders. One absolute path exists — the author's own macOS Blender profile in docs/evidence/archimesh-doctor/doctor-default.json — a personal machine path, no secret. The Python utilities make no outbound calls at install time and require no API keys; the Blender worker needs Blender 5.x and is optional.

Packaging notes (TalentDelta)

• Mixed licence: MIT repository plus a GPL-2.0-or-later Blender worker. The GPL files ship with source and licence text; the catalogue licence line must not read MIT alone.
• The 12 agents ship twice, unchanged: agents/ carries kind/max_turns/timeout_mins frontmatter, providers/claude/agents/ the same prompt body with maxTurns, so each harness gets native metadata. The 30 skills duplicate byte-for-byte under providers/claude/skills/.
• Jurisdiction-agnostic: no UK Building Regulations, Approved Documents, Part L/M or PAS 2035 text. Suitable for concept and coordination work, not statutory submission.
• Upstream release state is NOT_VERIFIED (v0.9.0); pypdf/ezdxf-based utilities are tested by the shipped test suite.