-
Notifications
You must be signed in to change notification settings - Fork 3
FOSS Quickstart
Status: Stable See also: Installation | Infrastructure-Setup | Core-Engine | Agents-and-Skills
Everything on this page is open source and runs on one machine. No private registry, no VPN, no credential store, no cloud account. The private addons add intelligence on top of this stack — they do not replace any of it.
| Service | Port | Role | Required? |
|---|---|---|---|
| hive-mcp | 7910 | the host itself, over nREPL | yes |
| Chroma | 8000 | vector store behind memory + semantic search | yes — the default :memory-store backend |
| Ollama | 11434 | local embeddings, local models | strongly recommended |
| clojure-lsp sidecar | — | code intelligence for Clojure projects | optional |
| NATS | 4222 | event bus for hivemind / swarm | optional, off by default |
| Prometheus · Grafana · Loki | 9090 · 3000 · 3100 | observability | optional |
Without Ollama the system still runs: memory stores and retrieves by tag, type and scope — only semantic search goes dark. That degradation is deliberate, and it's the same noop-fallback pattern the whole addon architecture uses.
git clone https://github.com/hive-agi/hive-mcp
cd hive-mcp
bin/hive-mcp-fossbin/hive-mcp-foss starts the FOSS services, waits until each is actually reachable
(compose reporting a container as running is not the same as its process accepting
connections), then execs the nREPL.
Then point your MCP client at it:
claude mcp add hive -- /path/to/hive-mcp/bin/hive-mcp-foss| Variable | Default | Effect |
|---|---|---|
HIVE_NATS=1 |
off | also start NATS (set :nats :enabled true in config too) |
HIVE_TELEMETRY=1 |
off | also start Prometheus, Grafana, Loki, promtail |
HIVE_SKIP_COMPOSE=1 |
off | services already run elsewhere — just wait and boot |
HIVE_NREPL_PORT |
7910 | nREPL port |
CHROMA_HOST / CHROMA_PORT
|
localhost / 8000 | point at a remote Chroma |
OLLAMA_HOST |
http://localhost:11434 | point at a remote Ollama |
WAIT_TIMEOUT |
90 | seconds to wait per service |
# 1. services
docker compose up -d chroma lsp-sidecar
# 2. embeddings
ollama pull nomic-embed-text
# 3. the host
clojure -M:dev:nreplCheck Chroma is alive: curl http://localhost:8000/api/v2/heartbeat.
The defaults are already the FOSS stack — {:services {:memory-store {:backend :chroma}}}
ships as the built-in default, so an empty config works. Override only what you move:
;; ~/.config/hive-mcp/config.edn
{:services
{:memory-store {:backend :chroma}
:nats {:enabled true :url "nats://localhost:4222"}}}Secrets are read from the environment, never committed. If you add hosted models later,
their keys live under :secrets — resolved from env vars at read time.
The two rituals are plain requests, not slash commands — the model reaches for the
project, memory, kg and kanban tools itself.
Start of a session:
hive
project workflow catchupusing pwd as dir
End of a session:
make memories on all learnings this session, kg connect them, sync kanban, and
workflow wrap
Between those, multi batches whole sequences into a single call — see the DSL verbs in
Tools-Reference.
| Layer | Licence | What it is |
|---|---|---|
| hive-mcp | AGPL-3.0 | the host — protocols, registries, orchestrators, server, memory CRUD, KG edges, swarm, session ritual |
| FOSS addons | MIT | lsp-mcp, basic-tools-mcp, clj-kondo-mcp, hive-connectors, bb-mcp, olympus-web-ui |
| Product addons | Proprietary | ranking, learning, emergence detection, domain intelligence |
Core ships a working noop default for every extension point, which is what makes the FOSS stack a complete system rather than a demo. Adding a private addon later changes no configuration you wrote here — it registers into seams that are already there.
Want to build your own addon into those seams? Creating-Addons scaffolds one in a single command.