Semstreams is the substrate. Everything built on it is either a product we run continuously on our own metal, or a lab where one capability gets pushed until the framework tells us where it breaks. Three more stand alongside: standalone services the ecosystem depends on but doesn't build on.
Semstreams has no synthetic benchmark suite standing in for real use. Every sem* project below is something we actually needed, built by driving one capability past comfortable.
That means framework gaps arrive as product pain rather than as a passing test. A scheduler problem found in SemBoids at 30Hz is a scheduler fix upstream. A projection-contract problem found by SemConnect against the OGC conformance suite is a contract fix upstream. The consumers are the design pressure, and the framework is what survives them.
Every processor in SemTeams comes from Semstreams. graph, rule, agentic-dispatch, agentic-loop, agentic-model, agentic-tools, all of it wired up in one call at startup. A complete multi-agent product, and nobody had to write the processing path.
That freed the Go in SemTeams to be product. Roughly 23,000 lines of CLI, sandbox runtime and HTTP middleware, none of it spent reimplementing graphs, rules or agent loops.
Extending it stayed declarative too. SemTeams adds 34 vocabulary predicates of its own, with no fork, no patched processor and nothing waiting on an upstream PR.
Want to check? Search the repo for a component factory registration. There is exactly one, and it belongs to the framework.
Stage is where a project sits in its own life. Tracks is the Semstreams release it pins, which is the honest measure of how close it runs to the current framework.
An older pin means the project is resting, not rotting. The labs get picked back up in rotation, and the pin is where they resume.
The framework, and the stack we run on our own metal every day. Core is a short list on purpose. These are the ones we get woken up for.
| Project | Description | Stage | Tracks | Stresses |
|---|---|---|---|---|
| semstreams | The framework. Streaming pipelines that produce a live semantic knowledge graph, reactive rules, workflows, and LLM agent orchestration. One binary, NATS the only dependency. | Active | v1.0.0-beta.160current release | the substrate |
| semsource | Ingestion engine. Turns the pile of code, docs, configs, URLs, and media in your projects into a live graph with stable IDs and provenance. Earns its keep when the answer depends on relationships, not filenames. | Betav1.0.0-beta.6 | beta.160current | ingest scale indexing profiles |
| semteams | Agentic team coordination. The product shell (UI, personas, rules, configs and the tooling around them) for a full multi-agent system whose every processor comes from the framework. | Active | beta.159recent | agentic loop framework completeness |
| semdev | GitHub issue in, clean-room-verified pull request out. Two human gates, harness-owned measurement, budget-bounded and unattended in between. No bespoke lifecycle state machine. | Active | beta.160current | agentic verification rigor |
| semmem | Institutional memory. Lessons, decisions and the evidence behind them held as graph entities instead of in each agent's private notes. A problem is filed with its proposed solution, and a trend becomes doctrine that later runs are handed automatically. The existing repo was built for semantic issue tracking and is being repointed at this, so read the description as the target rather than today. | Planning | unpinnedpivot pending | institutional memory lesson corpus |
| semstreams-ui | Operational glass. Graph explorer, flow builder and runtime monitoring for any Semstreams app. Backend-agnostic, discovering components, schemas and graph data at runtime. | BetaTypeScript | API-levelno Go pin | runtime API surface |
Each one exists to push a single capability past comfortable, against something real, before it is anyone's dependency. Some are in active rotation; the paused ones are picked back up deliberately, and their pin is where they resume. Some stay labs as long as there is another dial to turn. Some earn promotion into Core.
| Project | Description | Stage | Tracks | Stresses |
|---|---|---|---|---|
| semconnect | OGC API Connected Systems gateway. Owns the OGC product bundle (OMS, SensorML, SWE Common, SOSA) over the framework's graph and projection contracts. External qualification at 137/137 passing. | Active | beta.160current | standards conformance projections |
| semboids | Reynolds boids, split-screen against their own live graph. Three steering rules produce emergent flocking; zone steering is JSON rules you toggle from the UI, and boids are born and culled through the graph itself. | Lab | beta.160current | throughput rules engine |
| semmachina | Fiction-first AI RPG. The world holds a truth independent of any story told about it, so narrative drift becomes detectable. Narration can be diffed against the graph and corrected. | Lab | beta.160current | agentic graph as authority |
| semlink | MAVLink companion mesh for vehicle-local state, local rules, and intermittent peer replication. BlueOS-compatible, not BlueOS-dependent. Grew out of the SemGCS ground-control demo. | Paused | beta.141lagging | edge & offline replication |
| semops | Data-fusion common operating picture. Owns the operator COP, feed adapters, canonical COP model, and scenario runner over the framework's projection and ownership rules. | Paused | beta.114lagging | data fusion geospatial |
| semdragon | Agentic workflow coordination modeled as a tabletop RPG. Work items are quests, quality gates are boss battles, and trust tiers are earned through demonstrated competence rather than declared up front. | Paused | beta.21pinned | coordination trust modeling |
These three carry no Semstreams dependency to pin, so they sit outside the Core and Labs split entirely. They are standalone services the ecosystem leans on, and two of them earn their keep well outside it.
| Project | Description | Stage | Role |
|---|---|---|---|
| semembed | Embedding service built on fastembed-rs. Multi-architecture without emulation, OpenAI-compatible, and a materially smaller footprint than the alternatives. Runs in production behind SemSource, and stands on its own for anyone who wants neural embeddings without the operational weight. | Production | embeddings |
| servicesim | Deterministic HTTP simulator for the Exa, Tavily, and Perplexity research APIs. Proves your client sent the correct vendor request, not merely that it got a response. Fails closed, never dials outward, and redacts credentials before anything is retained. MCP mocking and streaming support are in progress, aiming to retire the hand-rolled LLM mocks scattered across the ecosystem and replace them with one thing worth trusting. | Activev0.1.1 | test infrastructure |
| seminstruct | Instruct-tier inference images for chat completions, classification and summarization. llama.cpp with a curated GGUF baked in and an OpenAI-compatible API. Exists so instruct-tier work can run locally when the hardware in front of you is the constraint. | Alpha | local inference |
Pins verified against go.mod on 15 August 2026. Semstreams was at v1.0.0-beta.160.