Releases·Tool Release
Silica-core: "No-Model-in-the-Loop" MCP retrieval engine for coding agents cuts tool calls 28%
A developer building Silica-core, an open-source MIT-licensed retrieval engine exposed over MCP and CLI, describes its design constraint: the harness owns the loop. The tool returns only located evidence (file path, section, line window, mathematical signals) and never summarizes or decides for the model, a pattern the author reports cut tool calls by 28% with coding agents like Claude Code, Cursor, OpenHands, and Zed.
1 engagement1 sourceWed, Sep 9, 2026, 03:11 PM
AI Overview
- Silica-core (MIT, open-source) exposes BM25 section-level retrieval over MCP and CLI, returning file paths, line windows, and coverage/terms_absent signals
- Lexical BM25 beats local vector embeddings for docs, specs, and papers, avoiding latency, memory footprint, and exact-term blindness
- Atomic silica_read tool with line slices and schema-light outputs cut agent turns 22% and tool calls 28% across 144 benchmark runs on 5.5M tokens
- Server runs locally via uv with no API keys, cloud telemetry, or background embedding workers
Related
Social·Community
Dev builds Silica-core: hierarchical BM25 + coverage metrics over vector embeddings for agentic local doc retrieval
A developer building a local retrieval layer for coding agents describes abandoning the default dense-embedding RAG stack in favor of hierarchical BM25 with coverage metrics. The switch addresses slow local embedding generation on large corpora and missed exact keywords, function names, and technical terms. The post offers practitioners an alternative pattern for agentic local documentation retrieval.
A developer building a local retrieval layer for coding agents describes abandoning the default dense-embedding RAG stack in favor of hierarchical BM25 with coverage metrics. The switch addresses slow local embedding generation on large corpora and missed exact keywords, function names, and technical terms. The post offers practitioners an alternative pattern for agentic local documentation retrieval.
1 engagement1 sourceredditWed, Sep 9, 2026, 03:15 PM
Releases·Tool Release
What I learned making an evidence-first local search layer for coding agents
A developer released sift-light, a local search layer for coding agents designed to prevent agents from answering from plausible-but-unverified excerpts. It prioritizes exact search, keeps literal evidence from being displaced by semantic candidates, and supports a validation step that inspects retained source against the file on disk.
A developer released sift-light, a local search layer for coding agents designed to prevent agents from answering from plausible-but-unverified excerpts. It prioritizes exact search, keeps literal evidence from being displaced by semantic candidates, and supports a validation step that inspects retained source against the file on disk.
15 engagement1 sourceredditMon, Oct 12, 2026, 11:19 AM
Products·launched
Chinvat: local MCP labor hub for governed multi-model task delegation
Chinvat is a free, open-source local MCP labor hub for Windows that allows any MCP-capable coordinator (e.g., Claude Code, Claude Desktop, Codex, Cursor, Hermes) to delegate work to local models, remote specialist models, Windows itself, and communication/publishing channels. It provides a persistent job queue, artifacts, and a policy layer that governs what crosses the bridge, solving the problem of managing multiple AI models and tools from a single governed interface.
Chinvat is a free, open-source local MCP labor hub for Windows that allows any MCP-capable coordinator (e.g., Claude Code, Claude Desktop, Codex, Cursor, Hermes) to delegate work to local models, remote specialist models, Windows itself, and communication/publishing channels. It provides a persistent job queue, artifacts, and a policy layer that governs what crosses the bridge, solving the problem of managing multiple AI models and tools from a single governed interface.
5 engagement1 sourceredditSat, Jul 11, 2026, 05:08 PM
Products·building
AI coding agent creates real cloud infra behind dry-run + confirm
An MCP tool plugin for AI coding agents (Claude Code, Cursor) that provisions real cloud infrastructure (databases, auth, storage, hosting) on the user's cloud account. The agent creates projects, wires keys, and writes integration code, but requires a dry-run preview and explicit confirmation before any paid resource is created. Solves the trust and cost-control problem of letting AI directly manage cloud resources.
An MCP tool plugin for AI coding agents (Claude Code, Cursor) that provisions real cloud infrastructure (databases, auth, storage, hosting) on the user's cloud account. The agent creates projects, wires keys, and writes integration code, but requires a dry-run preview and explicit confirmation before any paid resource is created. Solves the trust and cost-control problem of letting AI directly manage cloud resources.
15 engagement1 sourceredditSun, Jul 12, 2026, 11:23 AM
Products·launched
Local-first, git-native memory layer for coding agents (Claude Code, Cursor, Copilot)
Fluree built a memory layer for coding agents that persists context across sessions using a local-first, git-native approach. It solves the problem of agents forgetting past decisions and repeating mistakes, targeting developers using Claude Code, Cursor, or Copilot via MCP.
Fluree built a memory layer for coding agents that persists context across sessions using a local-first, git-native approach. It solves the problem of agents forgetting past decisions and repeating mistakes, targeting developers using Claude Code, Cursor, or Copilot via MCP.
1 engagement1 sourceredditThu, Jul 16, 2026, 01:34 AM