Add MCP server tool filtering support to agents-js #164
Conversation
🦋 Changeset detectedLatest commit: b73a090 The changes in this PR will be included in the next version bump. This PR includes changesets to release 5 packages
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
|
@seratch can you review please? |
seratch
added
enhancement
|
thanks for the comments, @seratch! esp. the point about moving filtering logic that makes sense for keeping the server implementation focused on transport. |
|
Thanks @seratch. That makes sense and i’ve removed the |
Co-authored-by: Kazuhiro Sera <[email protected]>
Co-authored-by: Kazuhiro Sera <[email protected]>
…sts and server impl
- Moved tool filtering logic to agent/runner layer - Removed server-side filtering and context coupling - Updated test suite to reflect new behavior
Reverts prior tool filtering interface changes and updates corresponding tests.
vrtnis
force-pushed
the
feat/mcp-tool-filtering-js
branch
from
4dce2d5 to
47ae347
Compare
* Handle function call messages with empty content * Resolve test failure * Create heavy-yaks-mate.md --------- Co-authored-by: Dominik Kundel <[email protected]>
There was a problem hiding this comment.
Did thorough review and I found your current implementation actually does nothing for allowing/blocking tools. I've implemented draft version of change here: seratch@73eb37a Hope sharing this clarifies things!
examples/mcp/tool-filter-example.ts
Outdated
| console.log('\nAttempting to write a file (should be blocked):'); | ||
| result = await run( | ||
| agent, | ||
| 'Create a file named test.txt with the text "hello"', |
There was a problem hiding this comment.
currently this is blocked either way because the mcp server has the access only to sample_files/
my diff changed this as well
There was a problem hiding this comment.
thanks for sharing your draft, @seratch! You’re absolutely right, my plumbing additions never actually invoked the filter, so allow/block lists weren’t doing anything.
I’ve since merged your implementation from 73eb37a and used that as the foundation. Appreciate the clarity your diff brought to this.
One question: should we also update the test to explicitly include toolUseBehavior: 'stop_on_first_tool' in the agent config? That setting ensures the run exits immediately with a "Tool \"write_file\" is blocked by MCP filter." error rather than continuing or looping, which might make the filter behavior clearer and more predictable in test output.
happy to add that if helpful!
…js' into feat/mcp-tool-filtering-js
Hey all - this PR brings a new tool filtering feature to our MCP integrations. This resolves #162 based on openai/openai-agents-python#861
You can now:
Use
createStaticToolFilterfor quick allow/block listsProvide a custom filter function to make dynamic decisions based on
runContextUpdated core API so
getAllToolsnow accepts yourrunContext, ensuring filters are applied correctlyDocumentation enhancements in
docs/guides/mcp.mdxNew example (
examples/mcp/tool-filter-example.ts) to demonstrate both static and dynamic filteringComprehensive tests covering static lists, callable filters, multi-server hierarchies, and cache interactions
All tests pass and you can try it out by
pnpm -F mcp start:tool-filter