Use Ostler from LangChain, LlamaIndex and AutoGen

Load Ostler's browser, research and form tools in LangChain, LlamaIndex or AutoGen over MCP. Tested code and the one-session rule for cite.

On this page

The one rule: keep one session open
LangChain
LlamaIndex
AutoGen
Other frameworks

Ostler is an MCP server, so any agent framework with an MCP adapter can use its browser, research and form tools. This page shows the three popular Python frameworks, with the code that was run and the output it produced.

What was tested, and what was not. Each snippet below was run against Ostler 0.2.0 and the bundled demo page (ostler fixtures): the adapter connects over stdio, loads the tools, and calls read then cite directly. No language model was in the loop, so the step that hands the tools to an agent is shown as documented usage and marked as not run.

The one rule: keep one session open#

cite only accepts a quote from a page that read fetched in the same Ostler session. A session is one running Ostler process. Adapters differ in whether they keep it:

framework	default behaviour	what to do
LangChain (`langchain-mcp-adapters`)	`client.get_tools()` starts a new Ostler process per tool call, so `cite` returns `NOT READ`	open `client.session("ostler")` and use `load_mcp_tools(session)`
LlamaIndex (`llama-index-tools-mcp`)	`BasicMCPClient` starts a new process per call, so `cite` returns `NOT READ`	open an `mcp.ClientSession` yourself and pass it to `McpToolSpec`
AutoGen (`autogen-ext[mcp]`)	`McpWorkbench` as a context manager keeps one session	use `async with McpWorkbench(params)`

Both failures were observed, not assumed: the NOT READ reply is NOT READ: <url> was not read in this session ... call read first. It is Ostler refusing to certify a quote from a page it did not read in that session, which is the point of cite. Run the agent inside the async with block; when the block exits, the Ostler process and its Chrome end and the read history is gone.

Every example assumes ostler is on your PATH (see install) and the demo pages are being served with ostler fixtures. --tools research pins the 21 research-profile tools; use --tools all for all 30 (profiles).

LangChain#

pip install langchain-mcp-adapters      # tested: 0.3.2, with mcp 1.30.0

import asyncio
from langchain_mcp_adapters.client import MultiServerMCPClient
from langchain_mcp_adapters.tools import load_mcp_tools

URL = "http://127.0.0.1:8787/packs/research/kestrel-lock.html"
client = MultiServerMCPClient({"ostler": {"transport": "stdio", "command": "ostler",
                                          "args": ["serve", "--tools", "research"]}})

async def main():
    async with client.session("ostler") as session:  # one session = one Ostler process
        tools = {t.name: t for t in await load_mcp_tools(session)}
        print("loaded:", len(tools), list(tools)[:8])
        await tools["read"].ainvoke({"url": URL})
        for quote in ["It opened in 1887 and was designed by Marta Ilves",
                      "Tomas Reyes designed the Kestrel Lock in 1887"]:
            r = await tools["cite"].ainvoke({"url": URL, "quote": quote})
            print(r[0]["text"][:90].replace("\n", " "))

asyncio.run(main())

Output (the second quote is a decoy that is not on the page):

loaded: 21 ['navigate', 'snapshot', 'read', 'click', 'type', 'press', 'hover', 'scroll']
VERIFIED [1] <page_content untrusted="true">http://127.0.0.1:8787/packs/research/kestrel-l
NOT VERIFIED: these words are not on <page_content untrusted="true">The Kestrel Lock</page

Tool results come back as a list of content blocks, so the text is r[0]["text"]. To hand the tools to an agent, keep the agent call inside the same async with block. Not run (no model available here; langchain and a model package are extra installs):

from langchain.agents import create_agent
async with client.session("ostler") as session:
    agent = create_agent("anthropic:claude-sonnet-4-5", await load_mcp_tools(session))
    await agent.ainvoke({"messages": "Who designed the Kestrel Lock? Cite the exact quote."})

LlamaIndex#

pip install llama-index-tools-mcp       # tested: 0.6.0, with mcp 2.3.0

import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client
from llama_index.tools.mcp import McpToolSpec

URL = "http://127.0.0.1:8787/packs/research/kestrel-lock.html"
SERVER = StdioServerParameters(command="ostler", args=["serve", "--tools", "research"])

async def main():
    async with stdio_client(SERVER) as (r, w), ClientSession(r, w) as session:  # one Ostler process
        await session.initialize()
        tools = {t.metadata.name: t for t in await McpToolSpec(session).to_tool_list_async()}
        print("loaded:", len(tools), list(tools)[:8])
        await tools["read"].acall(url=URL)
        for quote in ["It opened in 1887 and was designed by Marta Ilves",
                      "Tomas Reyes designed the Kestrel Lock in 1887"]:
            out = await tools["cite"].acall(url=URL, quote=quote)
            print(out.raw_output.content[0].text[:90].replace("\n", " "))

asyncio.run(main())

Output:

loaded: 21 ['navigate', 'snapshot', 'read', 'click', 'type', 'press', 'hover', 'scroll']
VERIFIED [1] <page_content untrusted="true">http://127.0.0.1:8787/packs/research/kestrel-l
NOT VERIFIED: these words are not on <page_content untrusted="true">The Kestrel Lock</page

The clean text of a result is out.raw_output.content[0].text; out.content is the noisy str() of the whole result. Pass url to cite explicitly, since it is required. To use the tools in an agent, build a FunctionAgent(tools=list(tools.values()), llm=...) and run it inside the same session block. Not run (no model).

AutoGen#

pip install "autogen-ext[mcp]" autogen-agentchat "mcp<2"   # tested: autogen-ext 0.7.5, mcp 1.30.0

The mcp<2 pin is needed: with mcp 2.3.0 installed, from autogen_ext.tools.mcp import ... failed in testing with ImportError: cannot import name 'RequestContext' from 'mcp.shared.context'.

import asyncio
from autogen_ext.tools.mcp import McpWorkbench, StdioServerParams

URL = "http://127.0.0.1:8787/packs/research/kestrel-lock.html"
params = StdioServerParams(command="ostler", args=["serve", "--tools", "research"],
                           read_timeout_seconds=60)

async def main():
    async with McpWorkbench(params) as wb:  # one session = one Ostler process
        tools = await wb.list_tools()
        print("loaded:", len(tools), [t["name"] for t in tools[:8]])
        await wb.call_tool("read", {"url": URL})
        for quote in ["It opened in 1887 and was designed by Marta Ilves",
                      "Tomas Reyes designed the Kestrel Lock in 1887"]:
            r = await wb.call_tool("cite", {"url": URL, "quote": quote})
            print(r.to_text()[:90].replace("\n", " "))

asyncio.run(main())

Output:

loaded: 21 ['navigate', 'snapshot', 'read', 'click', 'type', 'press', 'hover', 'scroll']
VERIFIED [1] <page_content untrusted="true">http://127.0.0.1:8787/packs/research/kestrel-l
NOT VERIFIED: these words are not on <page_content untrusted="true">The Kestrel Lock</page

A NOT VERIFIED result comes back with is_error = True on the ToolResult rather than as an exception. AssistantAgent takes the workbench through its workbench parameter (the parameter exists; the model call was not run):

async with McpWorkbench(params) as wb:
    agent = AssistantAgent("researcher", model_client=model_client, workbench=wb)
    await agent.run(task="Read the Kestrel Lock page and cite who designed it.")

Other frameworks#

Anything that speaks MCP over stdio works the same way: launch ostler serve --tools research (or forms, or all), keep the process for the whole task, and give the model the tools. For Claude Code, Codex, Cursor and VS Code, see use with your agent. Not tested here: CrewAI, the OpenAI Agents SDK, Semantic Kernel, and non-Python frameworks.

View this page’s source on GitHub