v0.1

toolc v0.1

First tagged release of the toolc engine: an optimizing compiler for agent tool surfaces. Many MCP servers (or plain REST APIs) in, one compiled MCP endpoint out.

Highlights

Compiler. A capability-graph IR with immutable originals and reviewable overlays, content-hash versioned artifacts, and five independently toggleable passes: dead-tool elimination, LLM-assisted description rewriting with sibling and cross-server disambiguation, consolidation of near-duplicate tool families into deterministically routed facade tools (optionally across servers), hand-authored macro inlining, and a retrieval-based selection surface (search_tools / call_tool).

Frontends. Compile from live MCP introspection, from an OpenAPI spec (one tool per operation, constraints carried into schemas), or from a plain documentation URL: an LLM drafts the spec under never-invent rules and every drafted endpoint is probe-validated against the live API before compiling.

Runtime. A drop-in MCP gateway serving mirror or compiled surfaces, with full call logging (arguments, results, parent-call fan-out), serve-time result compaction that preserves identifiers and sourcing metadata (with an async mode for very large results), and credential injection that model arguments can never override.

Benchmark harness. Judged, ablated evaluation: 60 tasks across six categories, five conditions, run against two agent models (1,800 trials). Headline results, full data in docs/: compiled surfaces match the raw federation's accuracy for a frontier agent at 5x lower cost per completed task, and improve multi-source arbitration from 90% to 100%; the benefit is capability-gated, with smaller agents favoring direct surfaces. See the paper: https://toolc.dev/paper.pdf

Install

npm: @toolc/cli, @toolc/core, @toolc/gateway, @toolc/shared, @toolc/harness (all 0.0.1). Or clone and run directly; Docker and self-hosting instructions in SELF_HOSTING.md.

License

Apache-2.0. The hosted console at https://toolc.dev is proprietary and adds managed endpoints, observability with live tail, error-driven fix recommendations, and usage-based billing.