Squirrel

AI Coordination Primal for the ecoPrimals ecosystem.

License: scyBorg (AGPL-3.0-or-later + ORC + CC-BY-SA 4.0) | Build: GREEN | Tests: 1,539 passing (default features) | Edition: 2024 | ecoBin: 4.4 MB | Methods: 43 IPC

---

What is Squirrel?

Squirrel is a sovereign AI Model Context Protocol (MCP) service. It routes AI requests, manages context windows, coordinates multiple MCP servers, and provides vendor-agnostic model selection through runtime capability discovery.

Any OpenAI-compatible server, cloud API, or local model can plug in through the same interface. Squirrel discovers services at runtime — no hardcoded names, no compile-time coupling. Every port and endpoint is overridable via environment variables.

See ORIGIN.md for the full story of how Squirrel was built using constrained evolution.

Owns

• AI task routing and provider selection (cost, quality, latency)
• MCP protocol coordination
• Context window management (context.create / context.update / context.summarize)
• Human dignity evaluation on AI operations (discrimination, manipulation, oversight)
• Session management and configuration
• Capability registry (config/capability_registry.toml)
• Deploy graph (squirrel_deploy.toml)

Delegates (via capability discovery — no hardcoded primal knowledge)

• Auth and crypto to any primal providing security.* capabilities
• Data storage to any primal providing storage.* capabilities
• Service mesh / HTTP proxy to any primal providing network.* capabilities
• GPU compute to any primal providing compute.* capabilities

---

Quick Start

# Build (static musl binary — default target)
just build-ecobin

# Run (server mode — listens on Unix socket)
cargo run -p squirrel -- server

# Client (send a JSON-RPC call)
cargo run -p squirrel -- client --method health.liveness --params '{}'

# Test
cargo test --workspace --lib --tests

# Full CI gate (fmt + clippy + test + deny)
just ci

# Lint (zero warnings required)
just clippy

# Coverage
just coverage

Socket Path

$XDG_RUNTIME_DIR/biomeos/squirrel-${FAMILY_ID}.sock

Fallback: /run/user/<uid>/biomeos/squirrel.sock or /tmp/squirrel.sock.

Capability symlink: ai.sock → squirrel.sock (auto-created for capability-based discovery)

Auth Model

Squirrel does not expose auth.mode — it delegates all auth to the security capability provider (any primal advertising security.* capabilities). This is intentional: Squirrel is the AI coordination primal, not an auth server. TCP and UDS transports share the same JSON-RPC method surface; neither implements auth methods locally.

Method Gate (JH-0)

Pre-dispatch capability gate at crates/main/src/rpc/method_gate.rs. Ships in GateMode::Permissive (no behavioral change). Classifies every JSON-RPC method as Public (health, identity, capabilities, discovery, auth, provenance) or Protected (AI inference, tool execution, context management). Prepares CallerContext and ResourceEnvelope structures for JH-2 enforcement when BearDog ionic token verification ships.

Compute Delegation

Squirrel delegates compute workloads to the ecosystem compute capability provider via JSON-RPC IPC. Detection order: COMPUTE_SERVICE_ENDPOINT → COMPUTE_ENDPOINT → COMPUTE_SOCKET → capability discovery. The UniversalComputeAdapter resolves the provider socket via capability discovery and forwards compute.* operations over JSON-RPC 2.0.

Inference Provider Discovery

At startup, AiRouter discovers inference providers from multiple sources:

1. HTTP providers: AI_HTTP_PROVIDERS env + vendor API keys
2. Local AI: LOCAL_AI_ENDPOINT / OLLAMA_ENDPOINT / OLLAMA_URL → Ollama-compatible HTTP
3. Inference endpoints: INFERENCE_ENDPOINT / AI_INFERENCE_ENDPOINT → auto-registers a RemoteInferenceAdapter for neuralSpring or any inference primal (UDS or HTTP)
4. Socket hints: AI_PROVIDER_SOCKETS → comma-separated Unix socket paths
5. Socket scan: COMPUTE_SOCKET → tiered capability discovery

Runtime registration: any primal can call inference.register_provider to dynamically add itself. UDS inference calls use a 120-second read timeout by default (override via SQUIRREL_INFERENCE_TIMEOUT_SECS).

---

Architecture

TRUE PRIMAL: Self-knowledge only, discovers everything else at runtime.

Fitness:   1,539 tests passing (default features) | 616 `.rs` files | ~188k lines | zero Box<dyn Error> in prod

IPC:       JSON-RPC 2.0 over Unix sockets (default)
Binary:    tarpc with automatic protocol negotiation
TCP:       JSON-RPC 2.0 over TCP via `--port` + `--bind` (newline-delimited)
Transport: Unix sockets → Named pipes → TCP (automatic fallback)
Provider:  provider.register / provider.list / provider.deregister (spring registration)
Lifecycle: ecosystem lifecycle.register + ipc.register + 30s heartbeat
Niche:     niche.rs self-knowledge (capabilities, costs, dependencies, consumed)
Edition:   Rust 2024
ecoBin:    Pure Rust — zero C dependencies in default build

JSON-RPC health (ecosystem standard): health.check, health.liveness, and health.readiness are the canonical method names. The system.* names (for example system.ping) remain as backward-compatibility aliases only.

Capability-Based Discovery

let ai_services = ecosystem
    .find_services_by_capability("ai.inference")
    .await?;

Vendor-Agnostic AI

• Cloud: OpenAI, Anthropic, Gemini via API keys
• Local: Any OpenAI-compatible server (Ollama, llama.cpp, vLLM) via LOCAL_AI_ENDPOINT
• Hubs: HuggingFace, ModelScope via MODEL_HUB_CACHE_DIR
• Custom: Universal provider interface

---

Project Structure

squirrel/
├── crates/
│   ├── main/                  # Main library and binary
│   ├── core/
│   │   ├── mcp/              # MCP protocol + AI coordinator
│   │   ├── auth/             # Auth delegation (capability-based client)
│   │   ├── context/          # Context management + learning
│   │   └── interfaces/       # Core trait definitions (Plugin, DynPlugin, Context)
│   ├── config/               # Unified configuration
│   ├── tools/                # CLI, AI tools
│   ├── services/             # Command services
│   ├── universal-constants/  # Shared constants, primal identity, sys_info
│   ├── universal-error/      # Unified error types
│   └── universal-patterns/   # Transport, security, federation traits
├── specs/                    # Specifications
└── justfile                  # Build automation (just ci/test/clippy/coverage)

---

Degradation Behavior

When Squirrel is unavailable, downstream consumers degrade as follows:

Domain	Degradation	Severity
ai.* / inference.*	AI queries fail; consumers fall back to offline heuristics or cached responses	HIGH
tool.*	MCP tool routing unavailable; local tools still execute if consumer has them	MEDIUM
context.*	Context sessions unavailable; consumers operate stateless	LOW
capabilities.* / identity.get	Capability discovery fails; static configurations or cached responses used	LOW
graph.*	BYOB graph parsing unavailable; pre-validated graphs still deploy	LOW
provider.*	Spring registration queued; springs retry on reconnect	LOW

Standalone mode: Squirrel operates fully without other primals. AI routing degrades to local-only providers. Adapter operations return clear errors when no capability provider is discovered. No primal dependency is hard-gated.

Stadial Pairing

Downstream Partner	Integration Surface	Validation
esotericWebb	ai.query, tool.execute, context.* — agentic AI for game narratives	AI provider availability, tool routing
projectFOUNDATION	ai.query, inference.* — AI-assisted thread analysis	Inference endpoint discovery, model selection
neuralSpring	inference.register_provider — inference backend registration	Provider lifecycle, UDS timeout (120s)
all springs	capabilities.list, identity.get — discovery substrate	Canonical envelope shape compliance

---

Code Standards

• unsafe_code = "forbid" in workspace [lints.rust] — enforced across all 12 crates
• clippy::expect_used + clippy::unwrap_used = deny workspace-wide (test-only cfg_attr allows)
• #![warn(missing_docs)] on all library crates
• cargo clippy with pedantic + nursery + cargo lints — zero errors under -D warnings
• #[expect(reason)] over #[allow] — dead suppressions caught automatically
• cargo fmt — zero formatting violations
• Pure Rust: zero C dependencies in default build (ecoBin v3.0 compliant — sysinfo removed)
• Production files under 800 lines (test-only files may be larger)
• SPDX AGPL-3.0-or-later license header on all .rs files
• Edition 2024 across all 12 workspace crates
• tracing for structured logging (no println! in production code)
• Typed errors via thiserror; .context() on all key error paths
• Zero-copy patterns: Arc<str>, bytes::Bytes, Cow<str> on hot paths
• Capability-based discovery (no hardcoded primal names — CapabilityIdentifier replaces enum)
• Human dignity evaluation on AI operations (discrimination, manipulation, oversight checks)
• Property-based testing via proptest for serialization invariants
• Dev credentials env-only (no hardcoded secrets in source)

---

License

scyBorg — the ecoPrimals triple-copyleft framework:

Layer	License	Covers
Software	AGPL-3.0-or-later	All code, binaries, tools, infrastructure
Mechanics	ORC	Primal interaction protocols, spring deployment niches, ecosystem topology, constrained evolution methodology
Creative	CC-BY-SA 4.0	Documentation, papers, diagrams, specifications
Reserved	ORC Reserved Material	ecoPrimals branding, primal names, logos

Governed by three independent nonprofits. No single entity can revoke any layer.

Copyright (C) 2026 ecoPrimals Contributors