Srijika MCP Server

This stdio MCP server lets Codex and other MCP clients understand, validate, plan, and safely scaffold a CLI-first Srijika TSX project without Desktop Studio. It also coordinates immutable-source React and Next.js App Router migration into a distinct new Srijika target through the same canonical developer engine used by CLI and editor adapters.

Node.js >=22.13.0 is required. The npm package uses its declared runtime dependencies. The Codex plugin build is separate and self-contained, including the TypeScript parser, so an installed plugin does not depend on an ancestor repository node_modules directory.

npx -y @srijika/mcp-server --project /absolute/path/to/project

In a monorepo, run one server per declared application root. The generated root .mcp.json names them srijika-<project-id> and passes a different explicit --project value to each process. Aggregate discovery, architecture, and Vite/Next owner-test verification use srijika workspace CLI commands; an MCP server never guesses between application packages or widens its filesystem boundary to the monorepo root.

For migration, use the srijika_create_react_migration → scan → deterministic ownership plan → bounded immutable slice context → slice review → token-only apply → slice verify → global verify → finalize flow. The source is never written; equal, nested, overlapping, or symlink-aliased roots fail closed. Every applied slice follows canonical owner/role obligations, includes source-to-target traceability, and writes atomically inside the target. Review persists the exact bounded payload and binds it to source, target, plan, and slice hashes; apply accepts only the returned token. Review also supports expected-hash removal of obsolete generated starter files. Scan, plan, and status use compact opaque, query-bound paging, and status returns pending apply handles for interruption-safe resume. Each slice advances only after zero Srijika diagnostics, architecture, and engine-executed TypeScript and production build pass. Global verification executes typecheck/build/test and snapshot-signs the results; caller-authored statuses, receipts, routes, screenshots, artifact paths, and pass details are rejected. For route/visual parity, the engine runs source and target apps at distinct loopback ports. It prepares an isolated temporary source copy and target runtime, derives routes, captures fixed-viewport DOM/redirect/error and PNG evidence, applies fixed thresholds, and persists a snapshot-bound manifest. Verify accepts only target and optional includeInstall; finalize rebuilds the runtimes, reruns gates, and recaptures.

Browser parity requires Playwright and Chromium. This package declares Playwright as an optional dependency; run npx playwright install chromium if the engine reports a missing browser. Capture fails closed if either runtime is unavailable. Unsupported or ambiguous behavior remains blocking. Compatibility and adapter findings are planning hints only and cannot be reviewed or applied; generic runtime ignores, wrappers, copied legacy trees, source dependencies, open target graphs, and unowned target modules block completion. The read-only srijika://docs/react-migration resource is the machine-readable contract. The server does not claim arbitrary automatic rewriting or guaranteed zero context loss; Codex performs reviewed semantic slice analysis.

The create tool accepts optional framework: "next-app-router" to fail closed unless the source is a Next.js App Router application. Without that selector, the shared scanner detects React Vite/CRA or Next.js and returns the matching framework inventory. Next route/runtime metadata is returned from the same session used by CLI, VS Code, and Desktop Studio.

If slice context proves an inferred owner wrong, the bounded ownership-review tool can correct canonical owner kind/name/path/role only before the first slice review or apply. It recomputes the plan ID and SCC-atomic slices; it never accepts caller-chosen target filenames. Ownership review and slice context bind the exact current plan ID plus source and target snapshots, so values from separate plan reads cannot be mixed.

The code-project tools are bounded to --project, use the shared Srijika developer engine, refuse overwrite, and enforce the canonical Feature → Slot → Part contract. Structure plans also support named behavior-hook and store-slice expansions; the server derives owner-prefixed paths, reports every move/update, and applies the reviewed one-time plan atomically without mixed flat/folder gateways. Flat-to-folder migrations scan the complete bounded project TS/JS corpus so imports outside the ownership roots are rewired too; unsafe entries, symlinks, incomplete scans, stale sources, or budget failures abort before any write. The read-only srijika://docs/code-first-architecture resource also defines the strict shared composites accepted by every surface: shared-ui (pure UI plus optional Types), shared-widget (required UI + Connector and optional runtime chain), and shared-capability (no UI/Connector and at least one runtime layer). Shared code never imports Features, consumers use only public shared boundaries, freehand shared folders and noncanonical names are invalid, and runtime Shared owner dependencies must remain acyclic. The resource also publishes the SRIJIKA4116 direct-child-UI boundary, SRIJIKA4117 passive-Types boundary, SRIJIKA4118 framework-free Logic boundary, SRIJIKA4119 computed-module ban, SRIJIKA4120 unresolved-project-alias boundary, SRIJIKA4121 unresolved or outside-root relative/project-source import boundary, strict UI external-runtime ban, safe bounded custom architecture configuration, and emitted PROMOTE/SPLIT recommendation semantics. Types are consumed only through import type/export type and never contain or reference runtime values. If the project omits its architecture block, canonical defaults apply. An explicit block must declare exact profile feature-slot-part-v1; missing or unsupported profiles fail closed. Generated validation reloads current config, and VS Code/Studio exact-file previews use the canonical resolved planner rather than hardcoded paths. The complete config requires sourceOfTruth: "tsx" and a bounded entry ending in the resolved UI suffix. Its optional architecture block exposes all twelve root/directory/suffix overrides. Root tsconfig.json is parsed as JSONC; exact or slash-delimited terminal /* paths aliases are authoritative, while extends, nonempty references, and any explicit compilerOptions.baseUrl are rejected. Empty references are allowed; Vite-only aliases are unsupported unless declared in root paths. Configured suffixes are compared case-insensitively and cannot overlap by suffix. The configured entry remains a strict, counted UI source even outside the ownership roots. Config reads are capped at 64 KiB, tsconfig.json at 1 MiB, and a validation/migration corpus at 4,096 sources, 32,768 entries, 4,096 directories, depth 32, 4 MiB per source, and 24 MiB total. Studio live preview also follows the configured entry, UI suffix, and Connector suffix. CLI watch uses one filtered recursive project-root watcher, reloads config, observes config/tsconfig/entry/resolved roots and future-root ancestors, and stays alive through temporarily invalid configuration until recovery. The optional versioned adoption.ownership block adds bounded brownfield managed/include/exclude paths, recognized legacy UI/Connector/Hook directories, and explicit adopted owners. MCP inspection and checks return the same immutable coverage/move/rewire plan as the CLI; strictPassed covers adopted owners, while passed becomes true only for honest full-project adoption. The same inspection and check path exposes source-only framework adapter metadata for exact next/link, next/image, and explicitly registered local presentation components. It never loads their implementations. App Router server/client and serialization failures use stable SRIJIKA5001–SRIJIKA5005 diagnostics documented in docs/NEXT_FRAMEWORK_ADAPTER.md. When Srijika Studio is running, the same server also exposes its authenticated document, layout, preview, and history bridge tools.