AI Devtools Daily — Monday, July 27, 2026

A neutral decision tool that takes your traffic profile, latency needs, hardware/cloud budget, and a target open model, and answers whether you should self-host at all — computing the real memory footprint at each quantization, the break-even token volume against day-0 hosts, and a "you cannot fit this" verdict up front.

Who it’s for

Platform/infra leads tempted by "free" open weights (K3 and the giant MoEs after it) who can't quickly tell whether self-hosting beats a Together/Modal endpoint for their volume.

Why now

K3's ~1.4TB footprint makes the "open but unrunnable" gap concrete and dated. Distinct from Railyard (Jul 17, spot routing) and Weighthouse (Jul 8, compliant hosting): Ballast is the pre-commitment TCO-and-fit calculator.

First version

A model-and-hardware database (params, active experts, footprint per quant, KV-cache growth) plus a calculator ingesting token volume + latency SLA, outputting self-host TCO, break-even vs named hosts, and a fit/no-fit verdict.

What kills it

Hosts publish their own calculators — but a vendor's always concludes "rent from us." The neutral, cross-host model (including "don't self-host") is what no host builds against its own margin.

A framework plus managed runtime for the MCP Apps surface the 07-28 spec formalizes: author interactive server-rendered UI components once and serve them through any MCP client, with state, theming, and auth handled — the "Vercel for MCP Apps."

Who it’s for

Tool and SaaS builders who want their MCP server to render a real interface inside agent clients instead of dumping text; platform teams standardizing internal MCP Apps.

Why now

MCP Apps ship as a formal extension on July 28 — a brand-new surface with zero incumbents. Distinct from Switchboard (Jul 14, conformance/migration): Kiosk is the build+host stack for the new UI capability itself.

First version

A component SDK compiling to the MCP Apps render contract, a local dev server with mock-client preview, one-command hosting with per-app auth/telemetry, and a starter template library.

What kills it

A big client vendor ships an official framework — but the neutral, cross-client framework + hosting + template marketplace is the ecosystem play a single client won't prioritize.

A reliability layer for the new MCP Tasks extension: checkpoint, resume, retry, and observe agent work that now legitimately runs for hours, so a Task that dies mid-run resumes from its last durable step instead of restarting.

Who it’s for

Teams building agents that do genuinely long jobs (large migrations, batch analysis, multi-hour research) on MCP.

Why now

The 07-28 Tasks extension makes long-running work a first-class protocol concept overnight — and durability/checkpointing/observability for it don't exist. Distinct from Kiosk (#2, the UI surface): Relay is the reliability/runtime plane.

First version

A wrapper/runtime around MCP Tasks that persists step state to durable storage, exposes resume/retry/cancel, and streams a per-Task trace with token spend and progress; integrates with stateless-core routing so Tasks survive restarts.

What kills it

Frameworks add native durability — but exactly-once/partial-failure is hard, and most bolt it on shallowly. A focused, MCP-native durability+observability layer spanning servers is defensible.

An independent lab-in-a-box that, the moment a major open model's weights land, runs a standardized hallucination + safety + refusal battery and publishes a plain "nutrition label" you can read before deploying.

Who it’s for

Teams deciding whether to self-host or day-0-host a fresh open model into a real workflow, plus hosts/gateways wanting a credible third-party quality label.

Why now

K3's early drop with no hallucination disclosure, paired with instant hosted availability, creates an evidence vacuum at the deploy decision. Distinct from Rate Card (Jul 10, frontier-launch due diligence): Assay is open-weight-specific, hallucination/safety-focused, timed to weights hitting Hugging Face.

First version

An automated eval harness (grounded-QA hallucination, safety/jailbreak, false-refusal, tool-use fidelity) that fires on a new major release, plus a published dated label and a hosted runner for your own prompt distribution.

What kills it

Model makers publish their own cards — but self-reported safety numbers are exactly what buyers discount. The independent, standardized drop-day label is the trusted artifact.

A runtime guard that detects when a coding agent is spiraling on a bug it created — repeated failing edits to the same region, oscillating diffs, test-passes-then-breaks cycles — and hard-stops it with a "you're fixing your own mess, hand it back" alert.

Who it’s for

Eng and platform teams running autonomous/semi-autonomous coding agents whose token bills balloon when an agent gets stuck fixing regressions it just wrote.

Why now

The ~44%-of-tokens-on-self-made-bugs figure, alongside standing distrust of AI code, makes the spiral a named, quantified waste. Distinct from Payload (Jul 23, after-the-fact outcome attribution): Backspin is the real-time circuit breaker.

First version

An observer around the agent loop that fingerprints edit/test cycles, scores "spiral risk" (same-region churn, diff oscillation, net-negative test progress), and trips a configurable halt-and-escalate — plus a dashboard of where spirals concentrate.

What kills it

Harnesses add loop-detection — but vendors optimize to keep their agent running (and billing). An independent, cross-tool circuit breaker whose job is to stop unproductive loops is the position a harness won't build.