loading broccoloading broccoloading broccoabout
brocco is the runtime + dashboard for everyone who already lives in claude, chatgpt, and cursor and wants to stop tab-switching. here is the wedge, the surfaces, and the people we built it for.
The wedge
"The agent is the cheap part. The workflow is the moat. Brocco hands you both."BROCCO PRINCIPLE NO. 1
Brocco is not another orchestration framework. It is the runtime that wires Claude, GPT, and local models into your stack. Drop a tool factory describing how to talk to your CRM, your warehouse, your homemade dispatch system, and the agent uses it on the next run. That is the wedge, and we hand it to you.
→ drop a tool definition, the agent uses it next run.
surfaces
same 9 agents, same 13 tools, same audit log. pick the surface that fits where you already work.
multi-agent panes, broadcast mode
open-the-app, type a goal, watch n agents stream in parallel. live token meter, jsonl audit, share via url hash.
open dashboardrest + sse
post /api/v1/run with a bearer key (byok passthrough) and stream the agent loop into any language, any runtime.
get /api/v1/agentsinside claude desktop
every brocco agent registers as a tool inside claude desktop, cursor, or any mcp-compatible client. one config block.
mcp setupwire your stack in 30 lines
a python factory describing how to talk to your crm, warehouse, or homemade dispatch system. the agent uses it next run.
tool factory docsone-click workflows
pre-wired multi-agent goals: market research, launch day, customer deep-dive, content sprint. fork them in /app.
browse recipesjsonl, exportable
every prompt, every tool call, every result. one append-only file per run. greppable, diffable, your siem can ingest it.
security overviewwhy we built this
the constraints we ran into building agents in production, and the design decisions they forced. each is non-negotiable in brocco.
one agent good at everything is bad at everything. we ship 9 specialists with their own tool lists, system prompts, and memories.
a researcher does not write code. a coder does not draft cold emails.
every step of every run is appended to a single jsonl file. you can grep it, diff it, replay it, hand it to your security team.
no black-box autonomy. no surprise tool calls.
broadcast one prompt to n agents. they run concurrently in their own panes. each finishes when it finishes.
one cup of coffee, five outputs.
server-sent events from the model into the dashboard, with no polling and no spinners. live token chip in the header during every run.
you watch the agent think, not wait for it to finish.
prompt caching is on by default. 80% cache hit on repeat workflows. live token meter shows cost in dollars while you wait.
a $49/mo tier covers 2,000 runs. solo founders ship in the green.
on free tier, prompts go from your browser to your provider. on paid, our hosted runtime calls anthropic with zdr enabled by default.
byok is not a pricing trick. it is a security posture.
features
everything you need to ship agentic workflows you would actually trust with revenue.
Supervisor pattern with sub-agent delegation. Specialists for research, code, outreach, ops, each with its own tool set and memory.
File, shell, HTTP, search, your custom APIs. Each tool is a Python factory. Type signature in, the agent gets it on next run.
Every prompt, tool call, and result is appended to JSONL. Replay any run. Diff two runs. Compliance-ready by default.
Per-agent KV that survives across runs. The agent remembers what it learned yesterday, today, and uses it.
System prompts and tool definitions cached on Anthropic edge. Cuts cost ~80% on repeated workflows. No flag to flip.
Run on Hetzner, Vercel, your laptop. SOC 2 Type II in progress. BYOK. Encrypted at rest. Your data never trains a model.
Anthropic, OpenAI, OpenAI-compatible (Ollama, vLLM, OpenRouter, Groq). Your key, your tokens, your data path.
Server-sent events end-to-end. Watch each token, each tool call, each parallel agent. No polling, no spinners.
On paid plans, brocco calls Anthropic with ZDR enabled by default. We never log your prompts after the run completes.
who is it for
concrete, real, what-it-does-for-you-tuesday-morning. not abstract benefits.
Type one prompt at 9am: "qualify yesterday signups, draft 5 personalized intros, post a launch update to Slack." Three brocco agents fan out in parallel. Done by 9:08am.
Zaps break when input changes shape. Brocco agents read the data and decide. Connect Stripe, Notion, Slack, Postgres, describe in English, the agent runs on cron with full audit trail.
Drop a topic in Broadcast mode, pick 5 agents (researcher, analyst, outreach, designer, planner). Get a brief, 5 social posts, 3 logo concepts, and a launch plan. One prompt, one cup of coffee.
brocco runs your business while you sleep. 100 agent runs free, every month, forever. no card.