Vaelor vs Serena

Scroll the table sideways to see every column.

Timings come from one laptop (Intel Core i9-9880H, 16 threads, 16 GB), Serena on 2026-10-06 and Vaelor on 2026-10-07, with other programs running (load average 8 to 18). They are an order of magnitude, not a benchmark.

Python 3.11+, uv, Go with gopls on PATH. Check out v1.65.16 of the Vaelor repository, then set language_servers: [go] in the project’s .serena/project.yml (Serena creates it on first start). This walks callers one symbol per call, tests included, and writes every symbol it reaches with its distance:

import asyncio, json, sys
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

PROJECT, OUT = sys.argv[1:3]
rows = lambda txt: [(f, s["name_path"]) for f, kinds in json.loads(txt).items() for lst in kinds.values() for s in lst]

async def main():
    params = StdioServerParameters(command="uvx", args=["--from", "serena-agent==1.7.0", "serena", "start-mcp-server", "--project", PROJECT, "--context", "ide-assistant"])
    seen = {("internal/parser/parser.go", "ParseFile"): 0}
    frontier, calls = [("internal/parser/parser.go", "ParseFile")], 0
    async with stdio_client(params) as (r, w), ClientSession(r, w) as s:
        await s.initialize()
        for depth in range(1, 6):
            nxt = []
            for f, n in frontier:
                calls += 1
                res = await s.call_tool("find_referencing_symbols", {"name_path": n, "relative_path": f, "max_answer_chars": 2000000})
                for key in rows("".join(c.text for c in res.content if getattr(c, "text", None))):
                    if key not in seen:
                        seen[key] = depth
                        nxt.append(key)
            frontier = nxt
    json.dump({"calls": calls, "seen": [[f, n, d] for (f, n), d in seen.items()]}, open(OUT, "w"))
    print(calls, "calls,", len(seen) - 1, "symbols")

asyncio.run(main())

Run it with uv run --with mcp python walk.py /path/to/vaelor-v1.65.16 walk.json.

Build the engine at v1.72.8 (make build; Go 1.26+ and a C compiler), start it with MCP_PORT=18897 OUTPUT_DIR=/tmp/vout ./bin/vaelor, and call impact_analysis over MCP with the path of the v1.65.16 checkout and depth: 5. The inline response is condensed to counts; the full lists are in the file it names under OUTPUT_DIR.

import asyncio, sys
from mcp import ClientSession
try:
    from mcp.client.streamable_http import streamable_http_client as streamablehttp_client
except ImportError:  # the client was renamed between mcp releases
    from mcp.client.streamable_http import streamablehttp_client

async def main():
    async with streamablehttp_client("http://127.0.0.1:18897/mcp") as streams:
        async with ClientSession(streams[0], streams[1]) as s:
            await s.initialize()
            r = await s.call_tool("impact_analysis", {"repo": sys.argv[1], "symbol": "ParseFile", "depth": 5})
            print("".join(c.text for c in r.content if getattr(c, "text", None)))

asyncio.run(main())

The first call builds the index. On a checkout the Go build cache has not seen, it can return "tier":"basic"; repeat the call until the response says "tier":"enhanced". The drivers and raw outputs are in docs/measurements/serena-vs-vaelor-2026-10-06/.