- Python 74.5%
- Vue 19.8%
- TypeScript 4.3%
- Shell 0.8%
- CSS 0.5%
- Other 0.1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
deebf323a9
run.sh already had a HIVE_HOST knob defaulting to 127.0.0.1, so LAN access came down to the default. The built UI is same-origin and the Vite dev server proxies /api and /ws, so no frontend change was needed: useSocket.ts builds the socket from location.host and the API paths are relative. Also closed a hole that LAN exposure would have widened: CORS was allow_origins=["*"], which on a LAN-reachable port lets any page in a visitor's browser drive the swarm. It is now opt-in via HIVE_CORS_ORIGINS, empty by default. Nothing needs it in either normal mode. README documents the new default and that the API has no authentication: any host that can reach the port can read and write any file the server's user can. Verified: bound 0.0.0.0, index and /api/agents serve 200 over 192.168.1.145, and an Origin: http://evil.example request returns no Access-Control-Allow- Origin header. The websocket handshake and /api/system/status were not verified (the latter probes provider endpoints and was still pending). |
||
| hive | ||
| tests | ||
| web | ||
| .env.example | ||
| .gitignore | ||
| PLAN.md | ||
| README.md | ||
| requirements.txt | ||
| run.sh | ||
HIVE
A persistent, self-improving agent swarm that runs against your own OpenAI-compatible endpoint. No cloud dependency, no framework lock-in.
Agents are permanent. They own a directory of files — memory.md, skills/,
tools/, system.md — and they edit those files themselves. You chat with them
in the browser and watch every turn, tool call, and thought stream live.
Quick start
cp .env.example .env # set OPENAI_BASE_URL, OPENAI_API_KEY, DEFAULT_MODEL
./run.sh --boot # builds the UI, starts the swarm on :8710
Open http://127.0.0.1:8710, or <http://<your-lan-ip>:8710> from another
machine — run.sh binds 0.0.0.0 by default.
No authentication. Anything that can reach the port has full control of the swarm: it can read and write any file the server's user can, start agents, and message them. Bind
0.0.0.0only on a network you trust, or put a reverse proxy with auth in front. SetHIVE_HOST=127.0.0.1to go back to loopback.
What an agent can do
bash, python, read and write any file on the machine (their own, another
agent's, or anywhere else — there's no file sandbox), edit their own memory.md,
write and read their own skills, post to a shared board, spawn new agents, and
message you without blocking on a reply.
Everything an agent does is written to SQLite as it happens, so a crash costs the current token, never the conversation.
Layout
hive/ the kernel: config, db, registry, tools, llm, agent loop, server
web/ Vue 3 + Vite chat UI
agents/ one directory per agent — memory.md, skills/, tools/, work/
state/ SQLite (WAL)
tests/ kernel tests (offline, stubbed LLM) + live API + true E2E
run.sh supervisor: relaunches the server, catches failed boots
Tests
PYTHONPATH=$PWD python tests/test_kernel.py # offline, stubbed LLM
PYTHONPATH=$PWD python tests/test_api_live.py # needs ./run.sh running
PYTHONPATH=$PWD python tests/test_e2e_llm.py # real HTTP endpoint end-to-end
Configuration
| Variable | Default | Meaning |
|---|---|---|
OPENAI_BASE_URL |
http://127.0.0.1:8000/v1 |
your OpenAI-compatible endpoint |
OPENAI_API_KEY |
sk-local |
key for that endpoint |
DEFAULT_MODEL |
default |
model id used unless an agent overrides it |
HIVE_MAX_CONCURRENCY |
8 |
agents allowed to think at once |
HIVE_HOST |
0.0.0.0 |
bind address; 127.0.0.1 for loopback only |
HIVE_PORT |
8710 |
listen port |
HIVE_CORS_ORIGINS |
(empty) | extra browser origins allowed cross-origin; usually not needed |
HIVE_MAX_TURNS |
40 |
tool-loop cap per user message |
PRICE_IN_PER_MTOK / PRICE_OUT_PER_MTOK |
0 |
per-1M-token prices; 0 = tokens only |