The best open-source AI Infrastructure tools, curated: 100 in the registry, including DeepSeek DeepInfer, Flint (Recursive — open-weights local coding model), Headroom (context-compression layer for AI agents). Every tool is open-source; each entry lists license, self-hostability, and links.
DeepSeek DeepInfer· tool, 2026-10-21 Apache-2.0 inference engine open-sourced by DeepSeek for running its V4.1-Flash model offline on your own hardware — NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs — instead of calling a hosted inference API, built from the official deepseek-ai/DeepInfer repo.
Flint (Recursive — open-weights local coding model)· tool, 2026-10-21 Apache-2.0 open-weights coding model from Recursive (Richard Socher) — a ~40B-total mixture-of-experts (~2.5B active, 262K context) that runs locally via llama.cpp, Ollama, LM Studio, vLLM or SGLang, plus an Apache-2.0 companion inference engine, so a team can run an agentic coding model on its own hardware instead of a metered cloud coding API.
Headroom (context-compression layer for AI agents)· tool, 2026-10-21 Apache-2.0, self-hostable context-compression layer (Rust core; Python/TypeScript SDKs) that compresses tool outputs, logs, JSON, RAG chunks and files before they reach the LLM to cut input tokens, shipping as a library, proxy, agent-wrap and MCP server, so a lean team can cut agent token spend on tooling it owns instead of a metered context-management SaaS.
Ming-3.0 speech models (Ant Group / inclusionAI — 5 open MIT TTS/audio/ASR models)· tool, 2026-10-20 MIT-licensed, self-hostable family of five Ming-3.0 speech models from Ant Group's inclusionAI across TTS, audio and ASR (from a 0.5B dense TTS up to a ~126B/6.5B-active audio MoE), with open Safetensors weights and documented local serving, so a lean team can own text-to-speech, audio and transcription instead of renting metered speech APIs.
Xing4.0-29B-Intent (TeleAI — open-weights intent/extraction MoE)· tool, 2026-10-20 Apache-2.0, self-hostable 29B-total mixture-of-experts model (~4B active) from China Telecom's TeleAI for schema-guided intent recognition, slot filling, NER, and relation/event extraction into structured JSON, with open Safetensors/FP8/GGUF weights a lean team can run locally instead of renting a closed extraction API.
Harness² (Google Research — recursive agent harnessing)· tool, 2026-10-18 Apache-2.0, self-hostable framework from Google Research that improves the harness around a frozen model (its instructions, guardrails, memory, tool guidance, subagents, skills, scripts, plugins) via a Propose-Probe-Compose loop with no retraining, so a lean team can raise an agent's benchmark results without touching model weights or paying for fine-tuning.
Thesys OpenUI (Open Standard for Generative UI)· tool, 2026-10-16 MIT-licensed, self-hostable open standard for generative UI from Thesys: instead of hand-building every screen, an agent emits a typed UI spec that OpenUI renders into real interactive components (cards, sliders, buttons, forms), so a model can return a usable interface rather than a wall of text.
Clef (Cloudflare) — open-weight decision model· tool, 2026-10-15 Cloudflare's Clef and Clef-flash are Apache-2.0, open-weight decision models (a Qwen3-27B backbone that returns a typed choice, not generated text), a self-hostable alternative to the proprietary, US-hosted, paid Jev classifier — run them locally or on Workers AI with no data leaving your own infrastructure.
FLock THIS/THAT Model — deterministic decision classifier· tool, 2026-10-15 An MIT-licensed, self-hostable decision classifier from FLock.io and Oxford researchers that answers in a single forward pass with zero generated tokens, so unlike a text-generating LLM classifier it cannot parse-fail or drift; a 1.88B hybrid-attention model adapted from decider-2b, free to run locally.
AISI eval infrastructure (ControlArena + the Inspect family)· tool, 2026-10-14 The UK AI Security Institute's MIT-licensed, self-hostable eval and AI-control stack: ControlArena (control protocols on Inspect AI), plus the Transect capability-eval runner and Inspect Cyber, giving a lean team a free, owned alternative to paid eval platforms.
SwanLab· tool, 2026-10-13 SwanLab is an Apache-2.0 AI training-tracking and visualization tool that logs metrics, hyperparameters, and hardware data, compares experiments, and supports team collaboration — an open-source alternative to Weights & Biases offered as both a self-hosted deployment and a cloud service. Python.
Omnara· tool, 2026-10-13 Omnara is the open-source managed-agent platform — a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API. You bring the agent; Omnara runs it with hosting, sandboxes, durable state, permissions, and secrets, against any model via OpenAI or Anthropic endpoints (also OpenRouter, LiteLLM, Ollama). Self-host with Docker Compose or use Omnara Cloud free with your own keys. Ships a CLI, TypeScript SDK, and MCP server, reaching agents via Slack and Discord.
OpenBot· tool, 2026-10-13 Gives each AI agent its own isolated computer (browser, file workspace, tools) behind a policy gateway that reviews and audit-logs every action before it runs. Works with AG-UI agents (LangGraph, Mastra, CrewAI, Pydantic AI); model-agnostic and self-hostable.
FreeToken· tool, 2026-10-13 Edge-native Mixture-of-Experts inference engine that runs frontier open-weight models on consumer hardware by treating GPU/CPU/host-memory as one bandwidth-adaptive tier — the repo's framing is "datacenter-scale model serving on your desktop." Apache-2.0, desktop app + Python. The 35B-on-8GB-GPU and 2–4×-vs-Ollama figures are from the project's arXiv paper (2608.16157), not the current README.
Langfuse· tool, 2026-10-13 LLM engineering platform: observability, evals, prompt management, traces, and datasets — for diagnosing why a marketing/GTM agent made a bad recommendation, not just that it failed.
Milvus for DeepSeek Harness· tool, 2026-10-13 A read-only DeepSeek Harness (DSH) plugin that lets an agent inspect and search Milvus or Zilliz Cloud vector stores from chat: exact entity lookup, scalar queries, BM25 full-text, and dense+BM25 hybrid retrieval. Strictly read-only — no insert, delete, schema, index, or admin surface, enforced at the execution boundary rather than by prompt. Works with local Milvus Standalone or Zilliz Cloud; dense search uses DSH-managed embedding providers. Apache-2.0, free, self-hostable.
REA (Reverse Engineer Anything)· tool, 2026-10-12 MIT-licensed CLI and MCP server that gives AI coding agents reverse-engineering capability with no source access — native binaries, JS/Electron apps, .NET assemblies, Android APKs, websites, Go binaries, WASM, firmware, and EVM smart contracts — bridging to Ghidra, Hopper, and IDA, all analysis run locally.
Drex 1.5· tool, 2026-10-10 Open-weight 9B decision model from Nace.AI that scores structured yes/no, multiple-choice, or rating questions over a given state instead of generating text — ranked #1 under 10B parameters on the public Decision Index 0.3.1 benchmark, runnable locally via llama.cpp/Ollama.
ML Drift· tool, 2026-10-10 Google's open-source (Apache-2.0), cross-platform GPU-accelerated inference engine for on-device machine learning, succeeding the GPU backend inside TensorFlow Lite/LiteRT — runs AI/ML workloads on mobile, web, and desktop GPUs via OpenCL, Metal, WebGPU, and OpenGL ES.
Onyx· tool, 2026-10-10 Open-source (MIT) knowledge and context layer that connects to 50+ applications to index and surface organizational knowledge for a team and its AI agents — hybrid search, agentic RAG, and custom agents, self-hostable with the Community Edition.
Rivet· tool, 2026-10-04 Rivet is an open-source actor runtime for stateful workloads (AI agents, realtime apps, durable execution), used as a self-hostable alternative to Cloudflare Durable Objects: you run the open-source control plane as a Rust binary or container on your own infrastructure. Open source: yes (Apache-2.0); self-hostable. Free OSS core with an optional paid managed cloud. ~6.2k stars; created 2023; active (pushed 2026-10-04). Maker-stated 12 ms p50 cold starts, ~72 KB per actor.
Strata· tool, 2026-10-03 Strata is an open-source local inference engine that runs large open models (the maker demos a 125B-parameter model) on a single consumer GPU, exposing an OpenAI/Anthropic-compatible API on localhost. Open source: yes (MIT); self-hostable. Free OSS; ~8.2k stars; active (~628 commits). Needs an NVIDIA or AMD GPU with 12 GB or more.
Microsoft Agent Framework· tool, 2026-10-02 Microsoft Agent Framework is an open (MIT) framework for building, orchestrating, and deploying AI agents and multi-agent workflows, with first-class support for both Python and .NET. Open source: yes (MIT); self-hostable. Free OSS; ~13.9k stars; very active (~3,400+ commits) from microsoft.
Magnitude· tool, 2026-10-01 Magnitude is an open-source inference engine for agents that optimizes itself for your exact hardware — it compiles and tunes its own kernels on your device, so open models run (per the maker's benchmark) up to 2x faster than llama.cpp, on Apple Silicon, NVIDIA, AMD, or CPU. Open source: yes (Apache-2.0); self-hostable. Free OSS; ~6k stars; very active (~1,000+ commits).
Jeeves· tool, 2026-09-29 Jeeves is a reasoning-augmented Qwen3.5-9B decision classifier that produces calibrated yes/no/rating calls and exposes the reasoning behind each one. Open source: yes (MIT); self-hostable (download weights, local CUDA inference or Python SDK). Free; ~211 stars; maker PostHog.
Jeff· tool, 2026-09-29 Jeff is a family of small (0.8B–2B) zero-shot decision and classification models you embed directly in code for routing, moderation, and intent calls at ~22–28ms latency. Open source: yes (MIT code + Apache-2.0 model weights); self-hostable (local jeff-serve on CUDA or Apple MLX, weights pulled from HuggingFace, no external API). Free; ~952 stars; maker firelex.
OpenShell· tool, 2026-09-29 OpenShell is a kernel-level runtime sandbox that lets fleets of autonomous AI agents use files, packages, APIs, and credentials while enforcing security policy without granting unrestricted system access. Open source: yes (Apache-2.0); self-hostable (kernel-level runtime you run yourself). Free; ~10.3k stars; maker NVIDIA; supports Claude Code.
Sensored· tool, 2026-09-29 Sensored is a streaming-first PII/PHI redactor that sanitizes text before it reaches an LLM. Open source: yes (MIT); self-hostable (core regex + NER runs locally; an optional AI-confirmation step calls an external API and fails open). Free; ~1 star; maker atomicpages.
textsnap· tool, 2026-09-29 textsnap converts any image, screenshot, or webpage into plaintext locally via OCR, running fully offline after the first model download. Open source: yes (MIT project; models Apache-2.0); self-hostable (pip install, ONNX runtime, no GPU, no cloud, no API keys). Free; ~186 stars; maker kouhxp.
Hindsight· tool, 2026-09-28 Hindsight is an agent memory system exposing retain/recall/reflect over four memory networks — world, experience, opinion, observation — with per-bank MCP endpoints and 60+ integrations. Open source: yes (MIT); self-hostable (docker run; PostgreSQL/Oracle AI DB). Free OSS (optional hosted from vectorize.io); ~39.9k stars; active.
Jevpipe· tool, 2026-09-28 Jevpipe is a Unix filter that pipes arbitrary data into a cheap System-1 model and returns typed decisions — yes/no, multiple-choice routing, or scores — for bulk agent judgments. Open source: yes (Apache-2.0) client, but depends on a hosted model via OpenRouter. Usage-based (~$0.023 per 1,000 decisions); launched 2026-09-28; 1 star.
Kern Sandbox· tool, 2026-09-28 Kern Sandbox is a rootless, daemonless container runtime that runs LLM-generated code and agent tool calls in kernel-enforced isolation, callable from a CLI or Python/Node SDKs with MCP support. Open source: yes (Apache-2.0); self-hostable; pricing free. Single static binary; 427 stars; active.
SurfSense· tool, 2026-09-28 SurfSense is an air-gapped, open-source (Apache-2.0) NotebookLM alternative: an AI research app that parses, indexes and embeds your documents entirely on your own machine (local SQLite, no telemetry), then generates summaries, reports, slides, podcasts and study guides. SurfSense AI runs on local models or any OpenAI-compatible API, and ships an MCP server for connecting external sources.
HydraDB· tool, 2026-09-28 Open-source distributed graph database in Rust built on object storage — decouples storage from compute so data nodes and indexers scale independently. Neo4j-compatible via Bolt 5.x, with an HTTP query API supporting OpenCypher and GraphBLAS traversal. An infra / agent-memory / context layer.
Keel· tool, 2026-09-26 Keel is a gated Rust harness that runs AI coding-agent output through five sequential verification gates and emits cryptographically-linked evidence bundles requiring human sign-off before deploy. Open source: yes (Apache-2.0 per GitHub SPDX); self-hostable; pricing free. 2 stars, fresh (last push 2026-09-26) and unproven.
One API· tool, 2026-09-26 Self-hostable LLM API gateway (MIT) that unifies 30+ providers — OpenAI, Azure, Anthropic Claude, Gemini, DeepSeek, and major Chinese models — behind a single OpenAI-compatible endpoint. Manages and redistributes provider keys so clients just swap the base URL, and centralizes access control, quotas, and usage monitoring. Ships as a single binary or Docker; an open alternative to LiteLLM.
Pipecat· tool, 2026-09-26 Open-source Python framework for real-time voice agents, multimodal apps, and live AI — composable pipelines with swappable STT (22+), TTS (35+), any LLM, and WebRTC/WebSocket/Twilio/WhatsApp transports, plus multi-agent composition and sub-200ms latency. Maintained by Daily; self-host free under BSD-2-Clause or run on usage-priced Pipecat Cloud.
Ctxfw· tool, 2026-09-25 Ctxfw is an Apache-2.0, in-memory Tree-Sitter AST 'context firewall' — an MCP server that prunes what a coding agent sends to the model, trimming the code context down before it costs tokens. It is on PyPI and self-hostable. It is nascent (2 stars). Its headline '-72.4% token mass' figure is a self-reported, unbenchmarked project claim, not a measured result.
Zro CLI· tool, 2026-09-25 Zro CLI is an MIT-licensed command-line tool that runs your existing agent harnesses — Claude Code, Codex, OpenCode, Hermes — against open models, so you can point the coding harness you already use at an open-weight backend. It is self-hostable and early: 2 stars, near-zero traction, created 2026-07-22, last pushed 2026-09-24.
BoxHaven· tool, 2026-09-24 BoxHaven provides persistent remote dev machines — cloud VMs — for AI agents such as Claude Code and Codex, and is self-hostable. It is licensed AGPL-3.0, a strong/network copyleft license: network use of a modified version triggers source-sharing obligations, which matters for anyone running it as a hosted or SaaS offering. It is nascent (6 stars).
Docker Sandbox Kit Spec· news, 2026-09-24 Docker's Sandbox Kit Specification is a vendor-neutral spec (v3) that packages an AI agent's network rules, credentials, tools and volumes as a pinnable OCI image (a 'Kit') enforced by conforming runtimes. The Apache-2.0 spec is open (github.com/docker/sandbox-kit-spec); Docker Sandboxes is the first runtime and Docker plans to take it to the CNCF.
Anthropic announces Claude Opus 5.5, live on AWS, GCP and Azure· news, 2026-09-23 On Sep 22, 2026, Anthropic announced Claude Opus 5.5, its new flagship model, available across AWS, GCP and Azure. Published pricing is $4 per million input tokens and $20 per million output tokens, with prompt caching at $0.20 per million read and $5 per million write. Anthropic's claims of benchmark superiority, 'Fable-5.1-level on most work' and roughly 40% lower run cost than Opus 5 are vendor characterizations, not independently verified here.
OpenAI announces GPT-6 Sol and Luna, a two-tier model refresh· news, 2026-09-23 On Sep 22, 2026, OpenAI announced two GPT-6 models: Sol, aimed at complex and coding tasks at $2 per million input and $10 per million output tokens, and Luna, aimed at high-volume clerical work at $0.10 per million input and $0.50 per million output. Both are available in the API (gpt-6-sol / gpt-6-luna), Codex and ChatGPT Work. OpenAI's claims of 'half as many mistakes / Astra-level factuality' and pricing superiority over GPT-5.6 are vendor characterizations, not independently verified here.
Tenderness· tool, 2026-09-23 Apache-2.0 Python library that renders synthetic, deterministic documents from text and images via Cairo/Pango to produce labeled datasets for training vision-language and OCR models. It is a rendering library, not a hosted generator app — you call it in code to emit page images with known ground-truth, so the data-prep layer for a document-extraction model is reproducible instead of hand-collected. Early and small (7 stars), for teams needing controlled training data, not a turnkey product.
Drop· tool, 2026-09-22 Apache-2.0 rootless Linux sandbox that runs untrusted programs and agent-generated code in disposable, isolated home directories using user namespaces, with optional gVisor for a stronger boundary. Access is governed by a TOML file and nothing needs root or a container runtime, so you can execute a coding agent's output or a third-party dependency without giving it your real user account.
Strands harness-sdk· tool, 2026-09-22 Apache-2.0 SDK (Python and TypeScript) from the AWS-backed Strands Agents project for building AI-agent harnesses that run in-process, with no hosted control plane. It is model-agnostic — first-class Amazon Bedrock plus Anthropic, OpenAI and Gemini, and local models via Ollama — so an agent you build stays portable across providers and can run entirely on your own hardware instead of being tied to one vendor's runtime.
Google AX· tool, 2026-09-21 Apache-2.0 agentic orchestrator from Google (Go) for running autonomous agent workloads at scale. You declare an agent task in a YAML manifest, its workspace and its gateway, and AX sandboxes the task, wires up the workspace, fences its network, and runs it, on Kubernetes, through a control plane. It integrates with LLM providers you supply, so it handles the isolation and scale-out while you bring the model. Early observers describe it as Kubernetes for agent executions.
Laya· tool, 2026-09-19 Apache-2.0 local decision engine (Python, open weights) for classifying and scoring text without generating prose. It returns typed choices, ordinal scores and boolean probabilities in one forward pass, with presets for routing, guardrails, moderation and support triage. Three checkpoints cover English, multilingual and task-tuned work. The project reports 33 ms on a T4, but also discloses weak zero-shot performance on the base model and the need for domain calibration.
MiniMax Code· tool, 2026-09-19 MIT source preview of MiniMax's terminal coding agent (TypeScript). The published source covers an interactive TUI, a headless CLI for scripts and CI, and an ACP endpoint for compatible editors. It can use a MiniMax account or a bring-your-own provider through OpenAI-compatible or Anthropic-compatible APIs, including local models. Search, plugins, multimodal tools, MCP and permission controls sit in the same workflow. The repository does not include the separate desktop application's source.
Bumblebee· tool, 2026-09-16 Apache-2.0 Go tool from Perplexity that inventories the packages and editor extensions installed on a developer machine and checks them against catalogs of known-compromised software, so a supply-chain compromise gets caught on the workstation rather than in production. It is a single static binary with zero non-stdlib dependencies, installed with `go install` and run locally; the compromised-software catalogs ship with it.
Anomalib· tool, 2026-09-13 Apache-2.0 deep-learning library from the Open Edge Platform (Intel lineage) for visual anomaly detection in images and video: Lightning-based, ready-to-use model implementations from experimentation through deployment, with OpenVINO export for accelerated inference. Claims the largest public collection of anomaly-detection algorithms; the bundled Anomalib Studio web app is pre-release.
zg (zvec-grep)· tool, 2026-09-13 A local-first search tool for people and AI agents that combines ripgrep, BM25, and on-device vector search across code, docs, and data — indexing and search stay on the machine. Auto-configures agents (Claude Code, Codex, OpenCode) and reports fewer tool-calls/tokens on SWE-QA-Bench.
litelm· tool, 2026-09-12 Open-source Python library (MIT, 182★) that extracts the core of LiteLLM — model routing, message translation, streaming, tool use and embeddings across 19 providers — into ~2,900 lines with two dependencies (openai, httpx), keeping API compatibility with LiteLLM for drop-in migration.
Agent-Native Registry: a catalog machines can read, sync, and cite without scraping — the definition· news, 2026-09-11 An Agent-Native Registry is a catalog published machine-readable first: every entry is structured data with provenance, consumers sync changes by date instead of re-scraping, and nothing unverified is presented as fact. The GTM Stacker Registry is a working implementation: 321 OSI-license-verified open-source GTM tools across 24 categories.
Raggy· tool, 2026-09-11 Open-source CLI (MIT, Python) for retrieval-augmented Q&A over your local documents: hybrid retrieval (vector + BM25), cross-encoder reranking, and fully local embedding/indexing via Ollama — answer generation can stay local or use a remote API, so private GTM docs never have to leave your machine.
Rune· tool, 2026-09-11 Rune, a fast GPU-rendered keyboard-driven IDE (Go) that folds code, terminals, CLI tools, language intelligence, debugging and AI agents into one multi-workspace environment, just went open source under GPL-3.0 — a polished commercial-grade dev environment moving to a copyleft license.
OpenObserve· tool, 2026-09-10 Open-source, OpenTelemetry-native observability platform (Rust): unified logs, metrics and traces with a dedicated AI/LLM layer that tracks cost, tokens, latency percentiles and error rates across models, plus agent graphs, session traces and evaluation/quality scoring — the monitoring layer for a self-hosted agent stack, an alternative to hosted APM.
Cynative· tool, 2026-09-09 Open-source framework (Go) for building your own security agents: ships 45 built-in agents that examine AWS, GCP, Azure, Kubernetes, GitHub and GitLab for privilege escalation, public exposure and supply-chain risks using frontier models. Cloud-security-focused — adjacent to, not core, GTM.
MagicVault· tool, 2026-09-09 Open-source secrets layer (Rust) that lets agents use enrolled credentials without ever seeing them: it injects secrets into approved browser fields, HTTP requests or spawned processes while keeping them out of prompts, replies and audit logs. VERY EARLY.
Agno· tool, 2026-09-03 OSS alt to LangGraph; ~41632 stars; via ossinsight.io
DVC· tool, 2026-09-03 OSS alt to proprietary MLOps asset stores; ~15800 stars; complements git-lfs, ComfyUI workflow files; Closest fit for versioning large generated assets/models alongside Git. Branch is otherwise thin — no purpose-built gen-asset versioning product exists. No model.
Haystack· tool, 2026-09-03 Pure Apache; production-focused
Hermes Agent· tool, 2026-09-03 Nous Research self-improving agent; persistent memory + autonomous skill creation; multi-LLM-provider; ~228k stars (anomalously high — verify hype); CAVEAT: autonomous skill-creation = arbitrary code execution by design — sandbox before use
Inspect AI· tool, 2026-09-03 UK AISI; rigorous eval (verified)
MiroFish· tool, 2026-09-03 Multi-agent AI prediction engine: builds a digital-twin sandbox of thousands of agents to forecast trajectories from seed data; ~71k stars (verify velocity); AGPL network-copyleft
Ollama· tool, 2026-09-03 Easiest local runtime; don't expose port publicly
Ray· tool, 2026-09-03 OSS alt to proprietary GPU-orchestration platforms; ~43500 stars; complements Celery, SkyPilot; General distributed-compute engine; also serves the cost/GPU/model orchestration branch. No model.
SkyPilot· tool, 2026-09-03 OSS alt to proprietary GPU-cost platforms; ~10500 stars; complements Ray; Provisions cheapest GPUs across clouds with spot/failover — the strongest OSS fit for the cost/GPU branch (STRUCTURAL: most cost-optimization is commercial). No model.
Unsloth· tool, 2026-09-03 AGPL on Studio/CLI; fastest kernels
JSON Feed 1.1 emerges as the agent-preferred syndication format· news, 2026-09-03 Autonomous crawlers index JSON Feed endpoints more reliably than legacy XML because payloads parse natively without entity resolution, escaping, or namespace handling. Publishers extend it with underscore-prefixed keys to carry structured agent metadata.
SIE (Superlinked Inference Engine)· tool, 2026-09-01 Self-hosted, OpenAI-compatible inference server that runs 85–100+ open models behind one API (encode/score/extract/generate), loading and LRU-evicting models by traffic so one GPU serves a rotating set instead of one server per model. Plugs into Qdrant, Weaviate, Chroma, LanceDB, LangChain, LlamaIndex.
Apple 'Agent Seer' research: API and parameter design break agent tool-calling· news, 2026-09-01 Apple research reporting that a large share of agent tool-calling failures trace to API and parameter design — schema shapes, parameter naming and documentation — rather than model capability. Practical implication for GTM-agent builders: design tool surfaces for the model, and keep a human approval step on consequential actions.
Itsuki· tool, 2026-08-29 Open-source memory engine for AI agents over REST API + MCP: captures conversations/events into typed, versioned, reversible objects (facts, entities, relationships, events) accessible across 26+ tools, with an audit trail and JSON export.
Tokensift· tool, 2026-08-29 Open-source, tokenizer-level static linter for LLM prompts and payloads — flags token waste (UUIDs, repeated blocks, pretty-printed JSON, base64 blobs) with fix suggestions. Library + CLI, 20 built-in rules; exact counts for OpenAI, calibrated estimates for Claude.
GLM-5.3 is now open-weight· news, 2026-08-28 Zhipu's GLM-5.3 was released as an open-weight model on Hugging Face — a runnable-locally option for bulk GTM classification/enrichment/drafting versus paying per API call. Exact commercial license terms are NOT independently verified.
Accept: text/markdown — serve Markdown to AI agents via content negotiation· news, 2026-08-27 An open HTTP technique (not a SaaS, no vendor/lock-in): clients send Accept: text/markdown and the server returns Markdown from the same URL that still serves HTML to browsers — cutting token bloat for agent readers. Copy-paste recipes for Nginx, Caddy, Next.js, Astro, Django, Rails.
Quartermaster (QM)· tool, 2026-08-25 An open, multiplayer agent harness for organizations: every employee and Slack room gets a scoped workspace (separate memory, files, permissions, sandbox), with shared skills, admin-gated promotion, crons/background watches, and one-command deploy to Fly/AWS. Model-agnostic (Claude/Codex/OpenCode).