The MCP bridge behind Amplifier's Agency bundle
Reaching many walled-off enterprise services from one agent normally means writing bespoke plumbing for each. Amplifier's tool-mcp bridge is how it avoids that.
And this isn't theory — it already runs at scale.
Its own reference deployment is production-proven: 60 capabilities drawn from 5 MCP servers — 41 tools plus 19 prompts — all reachable from one agent.
So the pattern works. Why isn't every service wired this way already?
Without a shared bridge, every service brings its own bespoke code, its own auth, and its own glue — per server, over and over.
The bridge exists to erase exactly this cost. Here's how.
The bridge does multi-server orchestration: stdio servers are spawned as child processes via command + args; Streamable HTTP servers connect over HTTP.
Many connections — but how do they become one usable toolset?
command + argsEach tool is namespaced per server, so many servers merge into a single agent toolset with no name clashes.
Connected and composed — and still, no per-server code to write.
MCP support is delivered as behaviors/mcp.yaml (v0.2.0), added with the --app flag. Servers resolve from four config sources — priority inline > env > project > user — with nothing to hand-wire, and zero overhead when unconfigured.
Now put a real set of services on top of it.
On top of this bridge, Samuel Lee's amplifier-bundle-agency — established 2026-04-03 in 2 commits — packages 16 zero-config MCP servers for Microsoft internal services. One conversation, sixteen services, no per-server code.
And the walls collapse for any services that speak MCP.
The Agency bundle is one instance of a reusable pattern: the tool-mcp bridge connects many MCP servers over stdio or Streamable HTTP, namespaces their tools into a single toolset, and needs zero per-server code. Point it at your services — the walls collapse the same way.
Bridge: verified from source Agency: corpus snapshot 2026-06-25
Data as of: Agency corpus snapshot 2026-06-25 (module.md last_updated 2026-06-26); tool-mcp cloned repo HEAD b32b7ce, 2026-05-03.
Commands run:
git remote -v — confirms origin microsoft/amplifier-module-tool-mcp; grep version pyproject.toml → v0.2.2sed -n '208,275p' amplifier_module_tool_mcp/manager.py — multi-server capability registrationgrep -n 'stdio_client|command|args' client.py — stdio (command+args) and Streamable HTTP transportssed -n '47,60p' amplifier_module_tool_mcp/wrapper.py — tool naming mcp_{server}_{tool}cat behaviors/mcp.yaml & sed -n '14,90p' config.py — composable behavior, 4-source config priority inline > env > project > usergrep -n 'Production Proven|60 capabilities' README.md — 60 capabilities, 5 servers, 41 tools + 19 promptscat samueljklee__amplifier-bundle-agency-2026-06-25-changes.md — 16 zero-config servers, 2 commits, 2026-04-03, Samuel LeeContributors: Samuel Lee (samueljklee) — author of amplifier-bundle-agency, both 2026-04-03 commits. Marc Goodner — earliest cloned tool-mcp commits (2025-10-28). sadlilas — latest cloned tool-mcp commit (2026-05-03, transport hardening).
Gaps: The Agency repo was not accessible via gh (“Could not resolve to a Repository”); its live source could not be read. Agency facts come solely from the corpus snapshot documenting the initial 2026-04-03 state. Claims such as 226 tools, a two-agent 75+151 split, Entra auth, and an IcM→Kusto→ADO→Teams flow could not be verified and are deliberately omitted. The bridge mechanism is fully verified from the cloned microsoft/amplifier-module-tool-mcp repo.