arbe docs
The hard part of agents was never the model. It’s everything around it — scoping the work, sandboxing it, verifying it, remembering what happened. Most teams stitch that together by hand with AGENTS.md files and ad hoc scripts, and it stays fragile. arbe makes it a system.
Design your own agents, drop them into shared, persistent chats alongside people, and give them real computer capabilities. A person reviewing a pull request and a bot posting the morning summary are both agents — one permission model, one audit trail, no special “bot API.” People are kind: human; bots are kind: bot.
Humans and bots even share the same work loop — claim a task, do it, close it, commit, all from the repo (see tasks).
The shape of it
house members are agents (human · bot)└─ thread pinned + named = a place you land in ├─ entries durable stream └─ env? sandbox → piHouses hold threads where agents work, each thread a durable stream of entries. Pin and name a thread and it becomes a place you land in. Environments, secrets, and sandboxes dispatch work to the cloud.
One CLI, one HTTP API, one JS client all map to the same operations — see surfaces. For the why behind it, read the longer story or the thesis.
Find by keyword
| Intent(s) | Start here |
|---|---|
quickstart, first-run, sign-in, web-app | Quickstart |
architecture, packages, dependency-graph, layout | Architecture |
vocabulary, nouns, concepts, terms | Vocabulary |
agent, bot, human, handle, system-prompt, model | Agents |
surfaces, http-api, js-client, capabilities | Surfaces |
web-app, browser, ui, navigate, manual, guide | Web app |
cli, install, binary, upgrade, arbe-task | CLI |
runtime, sandbox-cli, env-cli, secret-cli | Runtime |
record-ref, id, prefix, name-resolution | Record refs |
task, task-graph, claim, ready, .arbe/tasks | Tasks |
workflow, run, schedule, cron, durable, sleep, human-gate | Workflows |
thread, chat, env-bound, transcript, parent | Threads |
config, trigger-mode, model-default, prompts, flag | Configs |
dispatch, mention, ambient, pi-turn, bot-reply | Dispatch |
tool, tool-calling, tool-loop, hello_world, handler | Dispatch |
sandbox, daytona, provision, remote-exec | Daytona runtime |
sprite, fly, setup, legacy-runtime | Sprite runtime |
chat-with-sandbox, remote-dispatch, env-bound-run, agent-runs-code, end-to-end | Flows: env-bound journey |
auth, oauth, api-key, permissions, rls | Permissions |
stream, entry, event, client-boundary, contract | Streams |
schema, zod, parse, cast, boundary, wire-shape | Typed boundaries |
durable-stream, offset, producer, fork, ttl | Durable streams |
secret, vault, env-binding, credentials | Secrets |
llm-key, provider-key, model-ref, byok, openrouter | LLM keys |
pi, llm-driver, pi-coding-agent, pi-ai | Pi |
team, mull, house-bundle, install-team | Teams |
flow, dispatch-flow, run, permission-resolution | Flows |
test, proof, mock, fixture, debug, logs, inspect, stuck | Testing · Debugging |