The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Toolc listing page.
The optimizing compiler for agent tool surfaces. Point it at your MCP servers, or just a REST API's OpenAPI spec, or even a docs page, and get back one endpoint that serves a smaller, smarter tool surface: cheaper for every agent that connects, more accurate on every task, with every call logged.
Connecting N MCP servers dumps every tool definition into the model's context, degrades tool selection, and forces improvised multi-hop call chains. toolc compiles the federation instead:
search_x/find_x/get_x)
merge into single facade tools with deterministic routingsearch_tools/call_tool) when you want maximum context savings131 tools across 5 live servers, 30 tasks × 3 trials, judged answers (full grid · consolidation run):
| condition | success | cost / completed task | tool-def context |
|---|---|---|---|
| raw (all 131 tools) | 96.7% | $0.2631 | ~24,360 tok |
| compiled + consolidate | 96.7% | $0.0411 (6.4× cheaper) | ~845 tok |
| consolidate, no selection | 100.0% | $0.1335 | ~14,506 tok |
Consolidation dominates both ends of the frontier: the selection surface is the cost lever, consolidation lifts accuracy everywhere. Every pass is a config toggle; pick the tradeoff per deployment.
A downstream can be an MCP server or a bare REST API:
See SELF_HOSTING.md for Docker deployment.
toolc.dev runs this same engine as a managed service: compiled HTTPS endpoints, a console with a playground and full-payload log explorer, docs-to-spec drafting, and error-driven fix recommendations.
packages/core: IR (Capability Graph), frontends (MCP, OpenAPI, docs synthesis), passes, emitpackages/gateway: runtime MCP server (mirror + compiled) with call logging and compactionpackages/cli: the toolc commandpackages/harness: benchmark runner, grading, reportspackages/shared: config schema, errorsThe engine (everything in this repository) is Apache-2.0, forever: anyone can compile and serve their own tool surfaces locally with the exact same compilation quality as the hosted product.