Repository navigation
mcp-slim: Zero-dependency MCP stdio proxy for AutoGen (-50% to -98% schema tokens, $ref inlining, SecretGuard) #8473
Replies: 6 comments 1 reply
|
Salyyy (@SalyyS1) The $ref inlining at the proxy layer is the key differentiator here. Most LLM parsers struggle with deep $defs chains, often forcing fallbacks to generic string types or failing validation entirely. By resolving these references before the payload hits the agent, you prevent the LLM from wasting context window capacity parsing schema definitions it never needs to execute. This isn't just about token count; it's about reducing the cognitive load on the model to interpret static metadata repeatedly across turns. The cumulative stats tracking for wire bytes effectively quantifies this overhead reduction, which is often invisible in standard benchmarks but critical for high-frequency agent loops. |
|
Dheiver Francisco Santos (@dheiver2) Spot on. In multi-agent AutoGen loops, unresolved In |
|
How do you preserve semantic hints when inlining $refs? ShippingAddress and BillingAddress could have identical fields but different meanings. If the reference name disappears, does the surrounding schema retain enough context to tell them apart? Keeping the original definition name as a short description might be worth comparing against full inlining: slightly more tokens, but potentially fewer tool-argument mistakes. Does your evaluation check argument accuracy alongside token savings? |
|
agnotiq Great catch on twin When a property name already carries the role (for example We just shipped context-aware
|
|
Hii Salyyy (@SalyyS1), This looks like a really useful addition to the MCP ecosystem. A few things that stood out to me:
I also explored the browser playground. Having a zero-install, client-side environment to test schema compression and secret redaction before integrating the proxy is a nice touch. Overall, this addresses some practical MCP limitations, particularly token overhead and cross-provider schema compatibility. Looking forward to seeing how the project evolves! If this answer solves your doubts, could you please mark it as resolved? If the issue still persists, feel free to share the updated error so we can investigate further. |
|
Krish Khinchi (@krishkhinchi) Thank you for the detailed feedback. Regarding multi-agent orchestration and repeated tool calls in AutoGen:
We are continuing to expand the documentation with an end-to-end AutoGen multi-agent recipe showing lazy discovery and intent pruning across orchestrated agent teams. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone,
When connecting
StdioServerParams/mcp_server_toolsin AutoGen (autogen-ext) to enterprise MCP servers, two bottlenecks frequently surface:tools/listschema catalog (often 8k-15k tokens) plus verbose JSON tool outputs.strict=True/$refschema errors:mcp_server_toolsor downstream OpenAI/Anthropic/Gemini validators fail when MCP servers emit$defs/$refpointer chains, rootanyOf, or nullable["string", "null"]type arrays.We built mcp-slim, a zero-dependency Python 3.10+ stdio proxy that sits transparently inside
StdioServerParams:Repository & benchmarks: https://github.com/SalyyS1/mcp-slim
If this helps your AutoGen MCP workflows, please consider starring SalyyS1/mcp-slim on GitHub!
All reactions