Plan. Verify. Execute.
With ArcGen.
Describe what you want to build. ArcGen interviews you—or surveys your existing codebase—produces a verified architecture contract and dependency-ordered task graph, then executes it wave by wave through sub-agents.
Plan, verify, execute, navigate artifacts and inspect runtime facts over one MCP server.
Claude Code, Cursor, OpenCode, Codex, Gemini CLI, Antigravity, Copilot & more.
42% of a 32k context window with an 18,000-token CI ceiling and on-demand pulls.
Verifier review, plan review and explicit user approval after the deterministic floor.
A deterministic workflow
for complex development.
Turn ideas into a verified architecture, resolve dependencies, and execute tasks in parallel with explicit control.
From idea to execution.
After the verifier, a plan review checks scope fit and minimalism. The user gate presents the create, reuse and omit artifact list, wave summary and ownership map, then asks for your explicit approval before writing source code.
Complex work.
Clearly structured.
Every task declares explicit dependencies, disjoint file ownership globs, and objective acceptance criteria before execution begins.
Holistic Verifier Gate Brief
depends_on: T-02, T-03
- Assemble pre-rendered evidence map, dependency graph, and coverage matrix
- Require zero extra tool calls from the verifier sub-agent
One server.
A complete architecture toolkit.
One archgen-mcp server over stdio gives the model eight deterministic tool domains for planning, verification, execution, artifact navigation and runtime facts. Setup and CodeGraph administration remain on the CLI or CI control plane. One implementation, two ways in, so the skill and installer never disagree.
archgen_plan
Validates architecture contracts, drives the deterministic floor, profiles evidence-based scope and assembles the verifier brief.
archgen_tasks
Resolves topological execution waves, isolates failed dependencies and builds self-contained worker contracts.
archgen_docs
Indexes plan artifacts and serves targeted section slices, so agents never read whole files.
archgen_registry
Reads and synchronizes the features registry and module boundaries kept in AGENTS.md.
archgen_skills
Matches tasks against the capability-skill registry and injects normative rules into worker contracts.
archgen_refs
Loads ArcGen protocol reference documents on demand, one section at a time.
archgen_help
Returns the complete action schema and error recovery guide for a single tool domain.
archgen_platform
Reports the read-only runtime facts of the current harness — skill paths, MCP config, sub-agent dispatch and index state.
Your harness.
Your workflow.
ArcGen separates judgment from deterministic output. The Agent Skill (skill/SKILL.md) holds workflow policy and gate depth; the MCP server owns every artifact schema, dispatch brief and capability table.
Claude Code
Native background sub-agents with git-worktree isolation
OpenCode
Subagent tool and parallel headless worker runs
Cursor
Native background sub-agents with scoped file ownership
Codex CLI
Model-mediated spawn with flattened waves
Gemini CLI
Subagents-as-tools, hub-and-spoke wave execution
Antigravity
Async sub-agent invocation with forked context isolation
VS Code + Copilot
Hot-reload MCP wiring with .github/skills/ and .agents/skills/
Windsurf
Project .windsurf/skills/ store with global stdio MCP wiring
Kiro
Hot-reload MCP wiring with explicit skill resources
Zed
context_servers integration with a flat .agents/skills/ store
Trae
Hot-reload MCP configuration with .trae/skills/ and .agents/skills/
Universal (AgentSkills.io)
Universal .agents/skills/ layout for any compatible harness
One canonical skill across every environment.
Install per project or globally. Requires Node.js ≥ 20 with zero external CLI dependencies.
Installs the canonical skill, wires archgen-mcp into every detected harness and ensures CodeGraph. Global harness wiring is the default; use --location-project for project-local MCP configuration.
npx archgen-skill initLinks the canonical skill into every detected harness directory; pass --copy for real copies instead of symlinks.
npx archgen-skill installAuditable POSIX installer with zero dependencies beyond coreutils; use --no-mcp --no-codegraph when the optional server bootstrap is unavailable offline.
./install.sh --initLess context.
More useful work.
Measured on every commit via npm run test:budget. CI fails if the always-on footprint ever exceeds 18,000 tokens. The current total is 13,663 tokens— 42% of a 32k window.
Protocol reference
In total: 20,751 tokens
one section (278 tokens)
Loaded one section at a time
Capability skill rules
In total: —
injected in full (~160 tokens)
Injected into the worker contract
Dispatch brief
In total: 4,069 tokens
only the dispatched brief
Pre-rendered evidence, graph and reference checks
Artifact template
In total: 3,071 tokens
only the scaffolded template
Scaffolded on demand
Bring structure to your next project.
Initialize ArcGen with one command. Requires Node.js ≥ 20; reload your harness when setup finishes.
npx archgen-skill initInstall ArcGen.
Run npx archgen-skill init in your project root to install the canonical skill, wire archgen-mcp into every detected harness and ensure CodeGraph.
Reload your coding harness.
Restart or reload your IDE or CLI harness so the eight model-facing archgen_* MCP tool domains become active in your session.
Describe your development task.
Ask naturally to generate architecture, survey an existing repository, investigate a root cause or start autonomous wave execution.
Review the generated architecture.
Inspect the architecture contract, ADRs, plan decisions, wave summary and disjoint file ownership map in .archgen/<slug>/.
Approve and execute.
Grant explicit approval at the user gate, then let up to three sub-agents execute independent tasks wave by wave with full traceability.