mcp-code-orchestrate

MCP Code Orchestrate

The 98.7% fan-out token-cut. When you would otherwise call the same MCP tool N times and let every raw result land in the context window, write one script that does the loop, filters/aggregates in code, and returns only the distilled answer. Anthropic's "Code execution with MCP" reports a single example dropping from ~150,000 tokens of tool chatter to ~2,000 (≈98.7% fewer) by moving the fan-out into code instead of the model loop.

The pattern in one breath

Direct tool-calling is right for one call: ask, get the answer, done. It is pathological for fan-out — "enumerate-then-filter" across the 39-server MCP fleet — because:

1. Every tool definition you load is paid for on every turn (catalogue tax).
2. Every tool result (a 4 KB JSON blob × N functions / molecules / papers) is dumped verbatim into context, and the model re-reads all of it each turn.
3. The model is doing a mechanical loop (iterate, test a predicate, collect hits) that a for loop does for free, deterministically, off-context.

The fix: the agent writes one Python script (run via the Bash tool) that calls the underlying provider — its native CLI/library when one exists, else the MCP server over stdio/HTTP — loops over all the items, applies the filter in code, and prints only the distilled result (a shortlist, a count, the 12 hits out of 4,000). Raw payloads stay on disk / in process memory; only the answer enters the conversation.

Decision rule (use this / don't)

How a script reaches a provider — three access modes

1. Native CLI or Python library (best — zero MCP overhead)

2. MCP server over stdio (one session, many calls)

Example: spawn once, loop many.

sess = MCPSession(["npx", "-y", "mcp-searxng@1.8.0"],

env={**os.environ, "SEARXNG_URL": "http://127.0.0.1:8890"})

hits = [sess.call("searxng_web_search", {"query": q}) for q in queries]

sess.close()

3. MCP server over HTTP (already-running servers)

Concrete recipes

(a) RE xref-walk — find the hot functions across a whole binary