Prefex Copyright (c) Day Two AI. All rights reserved. Prefex is distributed as a binary (Beta), provided as-is with no warranty. "Claude" and "Anthropic" are trademarks of Anthropic, used only to describe interoperability. Day Two AI is an Anthropic partner; Prefex is built independently, with no Anthropic participation or endorsement. ================================================================================ Algorithms distilled into native Go (clean-room, NO source code used) ================================================================================ The following projects inspired ideas that were re-authored from scratch in Go. No source code from them is included in or distributed with Prefex. These are courtesy credits, not license obligations: - RouteLLM (Apache-2.0) https://github.com/lm-sys/RouteLLM - LLMLingua / -2 (MIT) https://github.com/microsoft/LLMLingua - caveman-compression (MIT) https://github.com/wilpel/caveman-compression - RTK (Apache-2.0) https://github.com/rtk-ai/rtk - Mem0 (Apache-2.0) https://github.com/mem0ai/mem0 - MemGPT / Letta (Apache-2.0) https://github.com/letta-ai/letta - OpenHands (MIT, core) https://github.com/OpenHands/OpenHands - GraphRAG (MIT) https://github.com/microsoft/graphrag - vLLM (Apache-2.0) https://github.com/vllm-project/vllm - Ollama (MIT) https://github.com/ollama/ollama - headroom (Apache-2.0) https://github.com/headroomlabs-ai/headroom (algorithms reimplemented in Go: content-routed compaction, recover-on-demand CCR, output shaping. The headroom Kompress *model* is run via the optional sidecar below, not reimplemented.) - codeburn (MIT) https://github.com/getagentseal/codeburn (ideas: dead-MCP-tool detection thresholds, cache-aware waste pricing, the embedded-pricing-snapshot approach. Re-authored as Setup Health.) - Invariant Scanner (Apache-2.0) https://github.com/invariantlabs-ai/invariant (idea: typed violation taxonomy for agent tool-call traces. Re-authored as the live policy gateway's Rego policy bundle; theirs is post-hoc trace analysis, ours gates on the wire.) - DefenseClaw (Cisco) (Apache-2.0) https://github.com/cisco-ai-defense/defenseclaw (ideas: tool-result content scanning, MCP tool-description admission scanning, severity-tiered outcomes, observe-then-enforce workflow. Re-authored in Go + Rego on prefex's wire position; their fail-closed posture and YAML policy schema deliberately not adopted.) - Chain of Draft (paper) https://arxiv.org/abs/2502.18600 (idea: draft-then-refine prompting to cut reasoning tokens. Re-authored as the effort/verbosity steer appended after the last cache_control breakpoint, so the cached prefix is never disturbed.) - Switchyard (NVIDIA) (Apache-2.0) https://github.com/NVIDIA-NeMo/Switchyard (idea: a TYPED classifier taxonomy for routing decisions — task_type, complexity, reasoning_depth, risk — rather than a single score. Re-authored as the advisor's typed schema. Their Rust wire-format translation layer was reviewed and not ported; prefex-auto reuses prefex's own translation.) - llmfit (MIT) https://github.com/AlexsJones/llmfit (idea: scoring candidate local models against detected hardware by weighted fit — VRAM, RAM headroom, disk, GPU speed class. Re-authored as internal/modelfit against internal/hardware's detected profile.) - OmniRoute (MIT) https://github.com/diegosouzapw/OmniRoute (ideas: an evidence-tagged, dated catalogue of free-tier plans; rate limits learned from response headers per key; a per-key circuit breaker. Re-authored as the prefex-auto free-pool catalogue and quota learner in Go. Its consumer-login providers, client fingerprinting, price table and advisory-only ToS labels were deliberately not adopted.) - Personal AI Router (NVIDIA) (Apache-2.0) https://github.com/NVIDIA/Personal-AI-Router (ideas: hysteresis bands on load thresholds; stale telemetry read as unknown, never zero; placing work on machines on your own network. Re-authored as prefex-auto's quota tracking and local-server capacity.) - Experiential (Apache-2.0) https://github.com/experientiallabs/experiential (ideas: paired-difference estimation with a variance floor — a constant sample cannot report zero uncertainty; and coerce-with-disclosure, where every request substitution is named as path->effective, logged and counted rather than applied silently. Both re-authored in Go. Their kNN-over-embeddings routing was reviewed and deliberately not adopted.) - GitAgentProtocol (opengap) (MIT) https://github.com/open-gitagent/opengap (idea: an agent declared as files in its repo, with a compliance block — risk tier, human in the loop, segregation of duties. Re-authored as prefex's agent-profile importer: the spec is read, and what it declares is enforced on the wire where prefex can, listed as unmapped where not.) - Token Meter (Splunk)(MIT) https://github.com/splunk/token-meter (ideas: evidence semantics, where every displayed number carries its own provenance — covered / partial / estimated / unavailable — and a period with no comparable prior renders "no comparable prior period" rather than a zero delta; a run-first information architecture that leads with the active session; and a single multi-mode execution chart in place of several single-purpose ones. Re-authored in Go and vanilla JS against our own on-wire data; their cost figures, runtime-adapter architecture and Git delivery page were reviewed and deliberately not adopted.) - agent-orchestration-standard (MIT) https://github.com/tomtomsatnav/agent-orchestration-standard (ideas: a file edited by two live agents within a window is a collision; a per-model-family silence threshold before an agent reads as stuck; "errored" only after three tool failures in a row; pausing background calls until a rejected plan window resets, else stepped backoff. Re-authored in Go as internal/agentwatch and the cache warmer's plan hold. Its live-graph web UI is separately included as source; see "Vendored source" below.) - codex (Seatbelt policy) (Apache-2.0) https://github.com/openai/codex (ideas: the deny-default macOS Seatbelt base policy, itself derived from Chrome's, and its sysctl allowlist. Re-authored in Go as the persona sandbox's profile generator, internal/sandbox.) - sandbox-runtime (Apache-2.0) https://github.com/anthropics/sandbox-runtime (ideas: Seatbelt rule ordering and log tagging, and the domain-allowlist egress proxy's matching rules. Re-authored in Go as internal/sandbox and internal/sandbox/egress.) ================================================================================ Third-party components downloaded on request (not bundled) ================================================================================ - Microsoft MXC wxc-exec.exe (MIT) https://github.com/microsoft/mxc Installed by `prefex persona setup` on Windows from Microsoft's own release, verified against a pinned SHA-256, and run unmodified with Microsoft's signature. Copyright (c) Microsoft Corporation. - apple/containerization (Apache-2.0) https://github.com/apple/containerization Built into prefex-vm, the helper for the macOS VM tier of personas; its structure follows the package's examples/sandboxy. Copyright (c) 2025-2026 Apple Inc. and the Containerization project authors. Its vminit image is pulled from the project's registry on first use. - a2aproject/a2a-go (Apache-2.0) https://github.com/a2aproject/a2a-go The A2A protocol types and server (a2a, a2asrv) behind the personas' inbox. Copyright (c) The A2A Authors. - Linux kernel from Kata Containers 3.26.0 (GPL-2.0, kernel; Apache-2.0, Kata) https://github.com/kata-containers/kata-containers Fetched by `prefex persona setup` on macOS for the VM tier, checked against a pinned SHA-256, and run unmodified. Source is available from the Kata Containers release. - Debian (bookworm-slim image; various free licences) https://www.debian.org The VM tier's default guest image, pulled by digest on a persona's first run; the browser preset installs Chromium from Debian's archive. ================================================================================ Third-party components bundled or invoked by the OPTIONAL Kompress sidecar ================================================================================ These apply ONLY if you opt into the text-compaction sidecar (compression.kompress: true). The sidecar runs out of process; the Prefex binary itself stays model-free. See deploy/kompress-sidecar/. headroom / Kompress (model + ONNX inference run by the sidecar) headroom's algorithms are reimplemented in Go (credited above); this entry is the keep/drop model the sidecar runs. License: Apache-2.0 https://github.com/headroomlabs-ai/headroom chopratejas/kompress-v2-base ModernBERT keep/drop model weights served by the sidecar (downloaded from Hugging Face at deploy time; not redistributed by Prefex). License: Apache-2.0 https://huggingface.co/chopratejas/kompress-v2-base ModernBERT (answerdotai/ModernBERT-base) The base encoder the kompress weights are fine-tuned from. License: Apache-2.0 https://github.com/AnswerDotAI/ModernBERT ONNX Runtime Inference runtime for the model (no PyTorch). License: MIT https://github.com/microsoft/onnxruntime Hugging Face Transformers Tokenization for the sidecar. License: Apache-2.0 https://github.com/huggingface/transformers ================================================================================ Third-party Go libraries compiled into the prefex binary ================================================================================ These are linked into the released binary and distributed with it: github.com/open-policy-agent/opa (Apache-2.0) embedded policy engine (policy gateway) github.com/redis/go-redis/v9 (BSD-2-Clause) optional Redis session store gopkg.in/yaml.v3 (MIT, Apache-2.0) config parsing modernc.org/sqlite (BSD-3-Clause) embedded dashboard database The complete set of third-party Go modules compiled into the binary (46 modules, direct and transitive — the growth over the earlier 22 is OPA's dependency tree) and their licenses is listed in THIRD_PARTY_LICENSES.md. Vendored source (third-party code carried in the Prefex source tree and compiled into the binary, rather than pulled as a module): github.com/gpustack/gguf-parser-go (MIT) third_party/gguf-parser-go/ GGUF model-file parsing, used to read local model metadata. The upstream LICENSE is retained in that directory. https://github.com/gpustack/gguf-parser-go agent-orchestration-standard web UI (MIT) internal/metrics/ui/live/aos/ Copyright (c) 2026 Tom Mitchell. The live agents graph at /dashboard/live (its JavaScript and CSS, unmodified), served in the dashboard page with its licence text embedded. The upstream LICENSE is retained in that directory. https://github.com/tomtomsatnav/agent-orchestration-standard Embedded data: LiteLLM model_prices_and_context_window.json (MIT) https://github.com/BerriAI/litellm — the model-pricing snapshot embedded in the binary and refreshed per release / via `prefex pricing refresh`. -------------------------------------------------------------------------------- A full copy of the Apache License 2.0 is available at: https://www.apache.org/licenses/LICENSE-2.0 If anything here is miscredited or mislabeled, please open an issue and we will correct it.