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

[dimslaev]

Topic: tool-calling (not covered by any of the 5 existing posts, all seeded 2026-07-13 under other topics)

Stories:

• https://github.com/Nanako0129/pilotfish — Fable 5 orchestrator + Sonnet/Haiku worker pattern for Claude Code
• https://github.com/dekart-xyz/geosql — Claude/Codex skill that writes PostGIS/BigQuery/Snowflake SQL with a map-in-the-loop
• https://aimaker.substack.com/p/claude-code-workflow-setup — wiring Codex into a Claude Code workflow via MCP

Why this one: all three are community/individual projects, not vendor products, and they add up to the same observation — the tool-calling glue between models (orchestration, review loops, domain skills) is being built by users on top of MCP/skills faster than Anthropic or OpenAI are shipping it themselves.

Written from feed summaries plus web search — source pages not read.

Per story, what the research added beyond the announcement:

• Pilotfish: the repo itself doesn't state the benchmark; multiple independent write-ups (Data Science Dojo, the-decoder, explainx) trace it to an Anthropic-published BrowseComp result (86.8% vs 90.8% accuracy, 96% performance at 46% cost) that pilotfish is packaging, not inventing. Search also surfaced the actual limitation — Fable's ability to notice unanticipated problems doesn't transfer to a cheaper worker model — which isn't mentioned in pilotfish's own docs.
• GeoSQL: the repo describes the map-in-the-loop mechanism (dry-run cost cap, geometry rendering) but not its own credibility problem — a Hacker News commenter flagged that the project's eval chart (single-digit % success) contradicts its prose (100% success claimed). Flagged in the post as a single-source, unresolved discrepancy.
• aimaker/Claude-Codex: the substack post itself is fairly abstract about MCP's actual role; a concrete community implementation (ching-kuo/claude-codex) supplied the specific mechanism — Codex returns only a structured verdict over MCP, capped at 3 rounds, to keep review cost bounded.

Single-source / disputed claims:

• GeoSQL's success-rate claim is disputed within its own sourcing (chart vs. prose) — called out explicitly rather than repeated as fact.
• Pilotfish's GitHub star count and exact role count came from directory-site listings, not the repo itself, so I left exact numbers out of the post and kept the claims qualitative.

Candidates dropped:

• https://www.anthropic.com/news/claude-sonnet-5 — already cited in the existing model-releases post (dedup).
• https://aidisruption.ai/p/anthropic-releases-sonnet-5 — same Sonnet 5 launch story as above, would double-cover ground the existing post already covers.
• Lower-scored tool-calling candidates (Grok UI-bug-fix agent, OfficeCLI, Enola, Julius/Browserbase, Claude Cowork cloud scheduling, FableCut) were all below the top 3 by score and didn't add enough beyond their announcements to bump ahead.

---

Generated by Claude Code