Get started

Plug your agent in, in one line.

Cascadian is one MCP server in front of one Postgres. Point any MCP client at it and the tools register themselves — your agent can call get_context_pack() on its next turn. A REST mirror is planned; the alpha is MCP-only.

01 · Connect

Add the server to your client.

ONE COMMAND

$ claude mcp add cascadian \
    --transport http \
    https://mcp.cascadian.ai/alpha

Add -s user to install for your account, or -s project to commit it to ./.mcp.json for the whole team. Tools register on connect — no SDK, no pipeline.

Then ask Claude“What happens to win rate and margin if we cut the discount tier?”→ it calls get_context_pack() and simulate_intervention() on its own · win_rate −2.1pp · gross_margin +3.4pp

02 · Headless agents

Wire it into an autonomous agent.

For an agent running without a human — a monitoring loop, a research pod, a fleet — skip the interactive client and connect the MCP server directly with a scoped key. The agent gets the same tools, bound to one workspace and attributed on every call. A Python SDK is planned; the alpha speaks MCP.

agent runtime · pythonSDK · PLANNED — ALPHA IS MCP-ONLY

from cascadian import Cascadian

# one scoped key per agent — bound to a workspace, attributed per call
cx = Cascadian(api_key=os.environ["CASCADIAN_API_KEY"])

# the agent assembles context, then simulates before recommending
pack = cx.get_context_pack(question="Should we cut the discount tier?")
ripple = cx.simulate_intervention(intervention={"discount_tier": "removed"})
print(ripple["downstream"][0])   # win_rate −2.1pp [±1.4]

Prefer to register the server with an SDK that speaks MCP (Claude Agent SDK, LangGraph, Vercel AI SDK)? Point it at https://mcp.cascadian.ai/alpha with the same scoped key and the tools register automatically.

03 · Verify

Confirm the tools registered.

Ask your client to list Cascadian’s tools — you should see the full surface, with the live server badge.

claude · /mcpEXPECTED OUTPUT

cascadian   ✓ connected — 27 tools
  bootstrap · get_context_pack · search_memory · …
  identify · simulate_intervention · record_outcome

04 · REST · planned

A REST mirror is planned.

Every tool will be mirrored over REST — same inputs, same JSON object back. The alpha is MCP-only; this is what the mirror will look like.

REST equivalentCURL · PLANNED

curl https://api.cascadian.ai/v1/context-pack \
  -H "Authorization: Bearer csc_live_…" \
  -d '{"question": "Should we cut the discount tier?"}'