Ideas,made tangible.

Weft keeps model protocols and capability dispatch in the core, while replaceable packages provide chat, an RSS reader, an AI video canvas, and multi-agent orchestration. packages/index.toml maps capabilities to packages; packages are classified as providers, foundations, or products, and can run as WASM, native, embedded, or service processes. Calls deny permissions by default, upgrades check permission changes, and 11 permission classes are enforced at call sites. Each package has an Ed25519 signature over its name, version, SHA-512 digest, and source. When a tool fails, its Closed/Open/HalfOpen circuit breaker controls whether calls continue. Local CPU tool routing uses ONNX INT8 cosine top-k without model-token use. The RSS reader first shortlists items with TF-IDF profiles, then asks an LLM to rank them. weft-core is published on crates.io, and the project provides a Windows installer; macOS and Linux releases have not been verified.

Direct code generation can produce output that runs and looks plausible in a diff, with errors that appear only during execution. CodeJev instead has the host enumerate AST-based code candidates and assign IDs; the model returns an ID or NONE, then the host assembles and verifies the result deterministically. In this comparison, which changed only the output format and used the same model, candidate selection succeeded in 5/5 cases and free generation in 0/5. In a separate selector comparison, the 350M model always returned the same ID, the local 1.5B model scored 2–3/4, Jev 4/4, and DeepSeek 5/5. These results describe small experiments only.

Separate identity, address discovery, and message history. The information-exchange service only provides account address/IP information; users can choose a real or relay IP, while each person runs or trusts a Store to retain, verify, and forward messages. The prototype connects nodes over WebSocket and Protobuf and uses a hash DAG to handle concurrent and out-of-order messages. Messages with missing parents wait until their parents arrive. Sorting uses DAG depth, local receive time, and hash; this does not guarantee the same total order across Stores. Signatures include parent references, and one-to-one Olm ciphertext has completed a round trip through real Stores. Group state, shared ordering, and new-device recovery rules are still being designed. Checks led to fixes in message signatures, ROOM_EVENT verification, device registration, and admission, but REST reads and inter-node WebSockets still default to no authentication or TLS. This is a research prototype, not a production product. Implementation is mainly AI-assisted; I make the design choices and assess the results.

RestoreWeave is self-hosted storage for NAS and files from multiple sources. It identifies content by full-file SHA-256 and logical length, so identical bytes share one stored copy while filenames and original paths remain distinct. Before writing, it shows a read-only plan with file count, logical size, estimated additional space, duplicates, and blockers; writing starts only after confirmation. Exports and restores recheck paths, lengths, and SHA-256. The index locates data but does not prove byte integrity; exact content can still be read and restored if the index or model is unavailable. Search supports keywords, structured conditions, tags, and notes. Semantic search requires explicitly installing and checking a local BGE model. This is an unreleased core preview with an API designed for loopback use.

GlassBox isolates agent execution by user and consists of nine application services. Containers drop all capabilities, add back only those needed, enable no-new-privileges, and use three network layers. Agent ports map only to temporary ports on 127.0.0.1. Idle sessions are reclaimed after 30 minutes, and orphaned containers are removed by label. Browser observation uses headless Xvfb and Chromium CDP without streaming the full desktop.

workflow-to-skill turns a goal and existing CLI, MCP, or HTTP tools into a reviewable SKILL.md; execution remains with the host’s existing tools and runtime. A real persistent session in DeepSeek Harness records one flow from generating a Skill to reusing it in a new session and obtaining a result. The repository keeps the related evidence map and use-case records.

Loom splits ideas into capability seams such as auth, CRUD, and tables, retrieves existing code from a personal component library, asks AI to select components, then has the client assemble a T3 project deterministically without calling an LLM. In one small experiment, all four assembled ideas were brought to a compilable state; all four from-scratch versions were stopped after repairs made them less coherent. The assembly runs used about one eighteenth as many output tokens. The experiment used only deepseek-chat and a small sample; “converged” means only zero TypeScript errors. loom-memory-mcp is published on PyPI.

Siftline separates research planning from search execution. Its Skill defines relationships to investigate, channels, queries, and stopping conditions. The CLI runs searches and returns schema-shaped JSON with provenance plus a machine-readable ledger. It performs no LLM inference, ranking, or conclusion-making; the caller assesses the returned material.

plate-redact masks plates in dashcam footage by combining results from multiple detection or recognition engines. Extra detections may leave extra solid blocks; a missed plate could expose it. The system predicts vehicle motion to apply masks early and uses solid fills. Its fail-closed publishing flow freezes input, takes an exclusive lease, publishes atomically, then decodes the output and checks whether vehicle tracks explain the masked regions. An unexplained region is marked UNKNOWN and blocks publication (test_fail_closed.py contains 17 checks). The repository includes no footage or model weights; third-party engines are loaded lazily when needed.

ModSpec uses .mspec.toml profiles to describe configuration and reusable .rule.toml files to define Hooks. A Rust CLI on the PC pairs with a phone agent over ADB. Validate checks the configuration, while verify checks device drift; changes apply only after both pass. The Hook scripting engine supports Frida-style before, after, and replace operations, plus structured events. It requires root access and LSPosed and remains a developer preview without a stable release.