content(blog): The tool-calling glue is coming from GitHub, not labs

frontend/content/blog/2026-07-14-the-tool-calling-glue-is-diy.md

@@ -0,0 +1,61 @@
+---
+title: "The tool-calling glue is coming from GitHub, not labs"
+description: "A cost-cutting Claude Code orchestrator, a geospatial SQL skill, and a Claude-Codex review loop, all built by individuals, not a model vendor's product team."
+slug: the-tool-calling-glue-is-diy
+topic: tool-calling
+date: 2026-07-14
+articles:
+  - https://github.com/Nanako0129/pilotfish
+  - https://github.com/dekart-xyz/geosql
+  - https://aimaker.substack.com/p/claude-code-workflow-setup
+---
+
+None of this week's three most interesting tool-calling projects came out of
+a model vendor's blog. All three are solo or small-team efforts stitching
+Claude, Codex, and cheaper models together through skills, subagents, and
+MCP — the parts of the stack Anthropic and OpenAI ship as raw primitives but
+don't finish for you.
+
+[Pilotfish](https://github.com/Nanako0129/pilotfish) packages a pattern
+Anthropic benchmarked but never shipped as a product: let Fable 5 plan and
+review inside your main Claude Code session, and hand the actual execution to
+cheaper Sonnet or Haiku subagents. On BrowseComp, that combination lands at
+86.8% accuracy against Fable-alone's 90.8% — roughly 96% of the performance
+for 46% of the cost. Pilotfish's contribution is packaging that as a handful
+of fixed roles and a one-prompt installer, rather than a sprawling catalog of
+agents. The catch shows up more in Anthropic's own write-up than in
+pilotfish's docs: Fable's edge isn't the checklist work you can hand off,
+it's noticing something's wrong when nothing on the checklist says so, and
+that doesn't delegate to a cheaper model. If your Claude Code bill is the
+actual problem, try it. If you're leaning on Fable for judgment calls, keep
+it in the loop instead of behind one.
+
+[GeosQL](https://github.com/dekart-xyz/geosql) turns Claude or Codex into a
+geospatial analyst that writes real PostGIS, BigQuery, and Snowflake SQL
+instead of guessing table names: it explores your warehouse schema first,
+dry-runs BigQuery queries against a 10 GiB cap before executing them, and
+renders results on a map so the agent can catch a geometry mistake a
+text-only review would miss. That map-in-the-loop step is the actual idea —
+the project's own comparison shows task success far lower without it. Worth
+flagging: a commenter on the project's Hacker News thread pointed out that
+the eval chart and the surrounding prose don't agree, a graph showing single
+digits next to text claiming success across the board. It's trending on
+GitHub regardless, and the underlying idea — let the agent see its output,
+not just describe it — generalizes well past geospatial data.
+
+[The AI Maker's write-up](https://aimaker.substack.com/p/claude-code-workflow-setup)
+on wiring Codex into a Claude Code workflow is less about one tool than about
+where MCP is actually landing a year in: as the connector to a handful of
+external systems, with skills doing the cheap, repeatable orchestration
+around them. A community project, ching-kuo's claude-codex, shows the shape
+concretely — Codex reviews Claude's diffs over MCP and hands back only a
+structured verdict, capped at three fix-and-recheck rounds, specifically so
+a second opinion doesn't cost more than the bug it catches. Small pattern,
+but it's the same one running through all three stories: nobody's waiting
+for Anthropic or OpenAI to ship the glue between their own tools, so people
+are writing it themselves in skills and thin MCP wrappers.
+
+Of the three, pilotfish is the one I'd actually install — the cost math
+holds up, and it doesn't ask me to trust a chart that contradicts its own
+text. GeoSQL's map-in-the-loop trick is the one I'd steal for something
+outside geospatial work, once someone reconciles those numbers.