Crossbar

Effortless local & self-hosted model backends for the Pi coding agent.

Crossbar is an extension for the Pi coding agent that makes wiring Pi to any local or self-hosted model backend effortless — zero hand-edited JSON, all setup inside Pi's TUI, with auto-discovery, multi-server support, live "currently loaded" indicators, and in-place model switching.

Built by Hypabolic.

Why Crossbar

Existing connectors stop short: manual config files, a single server at a time, Ollama-only discovery, or hardcoded model lists. Crossbar beats them on three axes:

1. Widest backend support — Ollama, LM Studio, llama.cpp, llama-swap, vLLM, OpenAI, Anthropic, plus a generic adapter for the OpenAI-compatible long tail (TabbyAPI, KoboldCpp, text-generation-webui, Jan, llamafile, …).
2. Easiest onboarding — /crossbar auto-discovers running servers on localhost and registers them. No JSON. API-key and no-auth endpoints, with a connection test before you commit.
3. Highest-fidelity UX — multiple simultaneous servers, a live loaded-model widget, and capability- driven model switching that gracefully hides what a backend can't do.

Supported backends

| Backend | Discover | Loaded indicator | Switch model | Load/Unload | Auth | |---|:--:|:--:|:--:|:--:|---| | Ollama | ✅ | ✅ live | ✅ implicit | ✅ | none (local) | | LM Studio | ✅ | ✅ live | ✅ JIT | ✅ | optional key | | llama.cpp (llama-server) | ✅ | ✅ (single) | ❌¹ | ❌ | optional key | | llama-swap | ✅ | ✅ live | ✅ proxy swap | ✅ | optional key | | vLLM | ✅ | last-known | ❌¹ | ❌ | optional key | | OpenAI | configured | — | pick model | — | API key | | Anthropic | configured | last-known | pick model | — | API key | | Generic OpenAI-compatible | ✅ (fallback) | last-known | ❌ | ❌ | optional key |

¹ Single model per instance. Run llama-swap in front of llama-server/vLLM to unlock switching — Crossbar detects it automatically and prefers it.

Full endpoint-level detail is in CAPABILITY-MATRIX.md; research notes and Pi-integration citations are in RESEARCH.md and ARCHITECTURE.md.

Install

# from npm
pi install npm:@hypabolic/crossbar

# or from git
pi install git:github.com/hypabolic/crossbar

# or a local checkout
pi install /path/to/crossbar

Then start Pi and run /crossbar.

Usage

• /crossbar (alias /local) — open the server manager. Discover/rescan servers, add one manually, enable/disable or remove saved servers, switch/load/unload models where supported, and inspect a server's currently loaded models. Esc/Back from a server returns to the server list.
• Adding a server registers its chat models immediately. Choosing Select model to use in Pi also makes that model current; Esc skips selection while keeping the server registered.
• Auto-discovery runs on session start: reachable no-auth servers on localhost are registered automatically; keyed servers are surfaced with a prompt to add them.
• Registered models appear in Pi's standard /model picker and /login provider list.
• The loaded-model widget shows what's currently resident (● live via introspection, ◷ with a (last-known) suffix where a backend can't report live state).

LAN discovery (beyond localhost) is off by default — no backend advertises over mDNS, so Crossbar never touches your network unless you opt in. Toggle it (and edit the LAN host / probe-port lists) from /crossbar → ⚙ Discovery settings; changes are saved to crossbar.json and apply on the next scan. When enabled with no explicit hosts, Crossbar auto-detects your machine's own private subnet (e.g. 10.0.0.0/24, derived at runtime from the system's network interfaces — never hardcoded) and sweeps it. You can override that by entering specific hosts, a CIDR (10.0.1.0/24), or a range (10.0.1.10-50); the expansion is capped at 1024 addresses. A full /24 sweep is a few seconds to ~20s depending on how many probePorts are set — narrow the port list to your backend's port to make it fast.

How it works

• Discovery probes localhost ports and fingerprints each server by response shape (e.g. Ollama's GET / banner, LM Studio's /api/v0/models state, vLLM's /version + owned_by), preferring the most specific match. The generic adapter is the low-confidence fallback.
• Registration maps discovered models onto Pi's built-in openai-completions / anthropic-messages providers via pi.registerProvider — Crossbar writes no streaming code for OpenAI-compatible servers.
• Two-phase startup: the async extension factory runs before Pi resolves models and preloads any enabled servers that have a persisted lastKnownModels catalogue (see src/preload.ts + registerCachedServer). This makes saved models available to --list-models, print/JSON/RPC, CLI --model selection, and --offline runs with zero network or UI. session_start still performs live refresh, auto-discovery, polling, and widget installation.
• Persistence keeps non-secret server metadata (including lastKnownModels) in ~/.pi/agent/crossbar.json; API keys live only in Pi's auth.json (mode 0600), keyed by the server's provider id. A brand-new server with no cached catalogue cannot be selected programmatically until an interactive /crossbar add or discovery persists its models (inherent cold-start limitation).

Security

• API keys are never written to crossbar.json, never logged, and never inlined into a provider config — Pi resolves them from auth.json at request time.
• Discovery is localhost-only by default.
• Crossbar adds no telemetry.

Development

npm install
npm run check   # tsc --noEmit
npm test        # vitest (conformance + unit + live-socket integration)

The BackendAdapter contract (src/core/) is the frozen boundary every adapter implements; the conformance suite (tests/conformance/) validates every adapter against it, and tests/integration/ exercises the real discovery path over live sockets.

CI

CI (.github/workflows/ci.yml) runs tsc --noEmit + the full test suite on every push and PR (Node 22 & 24).

License

MIT © Hypabolic