Switchyard

mcp
Security Audit
Warn
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 7 GitHub stars
Code Pass
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested
Purpose
This tool acts as a shared runtime layer that helps AI applications connect to user-provided credentials (like API keys via BYOK) and web sessions. It standardizes provider routing and diagnostics so developers do not have to rebuild these systems from scratch for every new AI product.

Security Assessment
The light code scan of 12 files found no dangerous patterns, hardcoded secrets, or requests for risky system permissions. However, because the core function is to route and manage end-user API credentials and active web sessions, the tool inherently handles highly sensitive authentication data. Developers must exercise caution regarding how these credentials are stored, processed, and logged at runtime. Overall risk is rated as Medium due to the sensitive nature of the data passing through the service, despite the clean automated scan.

Quality Assessment
The project is actively maintained, with its most recent push occurring today. It uses the permissive and standard MIT license and provides extensive, well-structured documentation alongside clear boundaries about its current capabilities. Community trust and visibility are currently very low. The repository has only 7 GitHub stars, indicating that it is an early-stage project that has not yet been widely peer-reviewed or battle-tested by the broader developer community.

Verdict
Use with caution — the codebase itself appears safe, but the project's low community adoption means it has not been broadly vetted, and its role handling sensitive credentials demands a careful manual review of your specific deployment configuration.
SUMMARY

Shared provider runtime for AI apps.

README.md

Switchyard

Shared provider runtime for AI apps.

Switchyard turns end-user BYOK + Web/Login access into one service-first
runtime that AI products can call without rebuilding provider routing,
credential/session handling, and diagnostics from scratch.

It is not a chat product, not a personal assistant, and not another
all-in-one AI platform. It exists to be the shared runtime layer that other AI
apps can depend on.

Switchyard social preview: shared provider runtime for AI apps. BYOK plus Web/Login. Read-only MCP descriptor. Runtime-diagnostics packet. Later lanes need fresh read-back.

Switchyard storefront loop preview.

Builder Thesis

One shared provider runtime for AI apps.

Use Switchyard when you want AI products to plug into real end-user access
lanes without every product re-inventing provider contracts, session logic,
and diagnostics.

Current Public Boundary

Today the truthful public story is:

  • Primary public front door: the GitHub Pages docs atlas plus this root
    README
  • Primary repo-native runtime surface: apps/service/ and the
    service/runtime HTTP contract
  • Secondary machine-readable surface:
    packages/surfaces/mcp/server.json, a read-only MCP descriptor
  • Builder-facing packet:
    distribution/claude-marketplace/plugins/switchyard-builder-suite/skills/runtime-diagnostics/
    plus starter packs and host examples
  • Not claimed today: official marketplace listings, official MCP Registry
    listing, npm publication, hosted multi-tenant runtime, write-capable MCP, or
    full consumer parity

Artifact-ready still does not mean listed-live.

Public Language Policy

Switchyard now treats the public front door as English-first.

  • The default landing path for global developers stays English-first.
  • Bilingual support remains available through glossary and i18n helper pages.
  • Live/browser realism notes belong in proof and runbook surfaces, not in the
    stable top-level product sentence.

Open The Right Door

If you need to... Open this first
understand the product in 30 seconds docs/media/30-second-overview.md
run the shortest first success docs/first-success.md
inspect what is really proved today docs/public-proof-pack.md
see what is package-ready vs listed-live docs/public-distribution-ledger.md
bootstrap a local workstation or inspect workstation-bound reality docs/runbooks/dev-bootstrap.md
browse the public docs front row docs/index.html and docs/README.md

If you are touching Switchyard for the first time, stop there.

The heavier shelves still exist, but they are not first-row front door pages:

  • support signboard and bootstrap runbook
  • contracts and blueprints
  • submission-packet accounting
  • builder catalog internals
  • testing/governance reference pages

The old Wave 1 working packs were relocated out of the public docs plane. The
legacy docs/blueprints/wave1/* paths now act only as relocation notes, while
the real working copies live under .agents/internal-docs/wave1/.

30-Second Version

If you only remember four lines, remember these:

  1. Switchyard is not another AI app.
  2. It is a shared provider runtime for AI apps.
  3. It turns BYOK + Web/Login access into a service-first substrate that other
    AI products can call.
  4. Today it ships a repo-native runtime, a partial read-only MCP surface, a
    runtime-diagnostics packet, starter packs, and truth-first public docs. npm,
    registry, marketplace, Docker, and broader publication remain later lanes.

First Success

If you want the fastest truthful first success, run the shortest bounded path:

  1. Start the local runtime:

    pnpm run start:service-local
    
  2. Prove the read-only truth surface is alive:

    pnpm run example:mcp-inspector
    
  3. Prove the runtime can accept one minimal invoke:

    pnpm run example:runtime-bridge
    

The default service port is http://127.0.0.1:4010.

If you want the full step-by-step path and failure routing, open
docs/first-success.md.

Why Switchyard Exists

Many AI products keep rebuilding the same messy layer:

  • provider contracts
  • auth/session plumbing
  • provider routing
  • diagnostics and remediation
  • builder starter surfaces

Switchyard exists so that this repeated work can become one reusable runtime:

a shared provider runtime that AI apps can plug into, instead of each app
re-inventing the provider layer alone

What It Is

  • a shared provider runtime for AI apps
  • a service-first runtime surface with SDK and MCP companion surfaces
  • a builder-facing repo that keeps proof, starter packs, and truth contracts
    aligned
  • a runtime layer that stays fail-closed on claims it cannot honestly prove

What It Is Not

  • not a chat product
  • not a personal assistant
  • not a control-plane-first SaaS
  • not a hosted multi-tenant runtime today
  • not a browser plugin
  • not a raw fork product of any upstream repo

V1 Scope

Switchyard V1 is intentionally narrow:

  • BYOK
  • Web/Login

Current Web/Login provider set:

  1. ChatGPT
  2. Gemini
  3. Claude
  4. Grok
  5. Qwen

Current BYOK code support must cover:

  • OpenAI
  • Anthropic
  • Grok / xAI
  • OpenRouter
  • Groq
  • Qwen API
  • Vertex AI
  • Bedrock

Explicit non-goals right now:

  • Agent Input Lane
  • Codex / Claude Code as supply-side sources
  • Gemini CLI
  • shared public credentials
  • multi-tenant account pooling
  • a hosted control plane

Architecture In One Sentence

Switchyard separates lane, provider, consumer, and surface so
that the runtime stays reusable even while builder routes and public claims stay
fail-closed.

High-level shape:

flowchart LR
    A["BYOK lane"] --> K["Switchyard kernel"]
    B["Web/Login lane"] --> K["Switchyard kernel"]
    K --> H["HTTP service surface"]
    K --> S["SDK surface"]
    H --> F["First-party integrations"]
    S --> D["Future AI app consumers"]
    H --> C["Future consumer compat"]

What Ships Now vs Later

Ships now

  • service-first runtime surface
  • partial read-only MCP surface
  • partial thin compat packages
  • runtime-diagnostics public skill packet
  • starter packs and host examples
  • proof-first docs atlas and public distribution ledger

Still later

  • official marketplace listings
  • official MCP Registry listing
  • npm publication read-back
  • Docker/runtime catalog publication
  • hosted multi-tenant runtime
  • full consumer parity
  • write-capable MCP

Proof And Reality Truth

Live/browser outcomes are important, but they are proof / runbook truth for a
credentialed workstation
, not the stable repo identity.

That means:

Use the proof pack when the question becomes:

  • what is really proved today
  • which blockers are external-only
  • which results depend on local credentials and browser session materials

Distribution Truth

Current distribution truth is intentionally narrow:

  • GitHub Pages storefront is live and remains the primary public homepage
  • repo materials are package-ready for the MCP surface and thin compat packages
  • official marketplace or registry publication is not claimed yet
  • builder packets and starter packs are public repo surfaces, not official
    listings
  • packet-scoped host receipts, including the
    switchyard-runtime-diagnostics packet, belong in the packet's own manifest
    and README; they do not upgrade repo-wide npm, marketplace, or official
    MCP Registry truth

See:

If you need the exact heavy-lane packet or older staging material, treat them as
deeper shelves rather than the default first stop.

Docs Atlas

First-row public routes

Deeper public shelves

Internal-only working packs

  • Wave 1 contract/evidence packs now live under .agents/internal-docs/wave1/.
  • The old docs/blueprints/wave1/* locations remain only as relocation notes so
    older links do not break.

Truth Rules

  • supported means there is a durable, repo-backed public surface now.
  • partial means a real narrow slice exists, but it is not the full promised
    shape.
  • planned means the route is intentional but not landed.
  • research means investigation exists but support does not.
  • not now means the surface is explicitly outside the current public front
    door.

Security And Local Runtime Boundaries

  • Do not publish cookie bundles, browser profiles, or other credential
    materials.

  • .runtime-cache/, .agents/, and .env* stay out of public release
    surfaces.

  • Repo-local cleanup only applies to Switchyard-owned runtime artifacts, never
    to machine-wide caches or other apps.

  • Before live/browser/cleanup actions, run:

    pnpm run scan:host-process-risks
    

For local runtime hygiene, browser/session acquisition, and operational
footprint, go to docs/runbooks/dev-bootstrap.md.

Verification Entry Points

  • pnpm run typecheck
  • pnpm run test:coverage
  • pnpm run test:docs-frontdoor
  • pnpm run build

One Final Sentence

Switchyard exists because AI apps should share one honest provider runtime
instead of each product rebuilding the provider layer alone.

Reviews (0)

No results found