Agent patterns¶

Recipes that combine Kitsune's tools. Tools marked full profile in the tools reference need KITSUNE_TOOLS=all.

Test the server you're building¶

The MCP REPL. Your agent runs, fixes and re-runs its own tool without anyone restarting the client:

connect("uv run python my_server.py", name="dev")
shapeshift("dev")
call("summarize", arguments={"url": "https://example.com"})  # wrong result

# … the agent edits the tool …

reload("dev")  # fresh code, tools remounted
call("summarize", arguments={"url": "https://example.com"})  # check the fix
release("dev")  # when done

Full walkthrough with a planted bug: hot-reload tutorial.

Check a server before you rely on it¶

Uses inspect and test from the full profile.

search("web scraping", limit=5)
inspect("firecrawl")  # tools, schemas, required credentials
test("firecrawl")  # quality score 0–100
shapeshift("firecrawl", tools=["scrape_url"])

Chain servers for a multi-step task¶

Mount one server at a time and drop it before the next:

shapeshift("brave", tools=["brave_web_search"])
brave_web_search(query="MCP server list")
shapeshift()

shapeshift("mcp-server-fetch")
fetch(url="https://example.com/mcp-servers")
shapeshift()

shapeshift("next-server") also replaces the current form on its own; the explicit shapeshift() keeps the transcript easier to follow.

Keep a stateful server alive¶

Servers that hold a device or a session (audio, browsers, databases) should run as a persistent connection, so every call reaches the same process:

connect("uvx voice-mode", name="voice", timeout=30)
call("speak", "voice", {"text": "Hello"})
call("speak", "voice", {"text": "World"})  # same process, device stays open
release("voice")

shapeshift() with no argument unmounts tools but does not stop a connect()ed process. Always release() it.

Let Kitsune pick the server¶

auto("current time in Tokyo")
auto("list issues on acme/api", server_hint="github")

auto() searches, picks a server, infers arguments and calls. Give a server_hint when you already know the server; it skips the search.

Benchmark before a batch¶

Uses bench from the full profile.

bench("mcp-server-time", "get_current_time", {"timezone": "UTC"}, iterations=3)
# p50 / p95 / min / max in ms

shapeshift("mcp-server-time")
for tz in ["UTC", "Asia/Tokyo", "Europe/Berlin"]:
    get_current_time(timezone=tz)
shapeshift()

Compare candidates¶

search("postgres", compare=True)  # token cost and credentials side by side

With the full profile, compare("postgres", limit=4) probes the candidates in parallel.