Blocks.ai claims CLI-based agent tools can avoid roughly 26,000 tokens of MCP schema overhead, making CLI access materially more context- and cost-efficient for large tool sets.

state: expiredheat: lowuncertainty: highconvergesscott: mediumagent-harnesses inference-economics mcp tool-useBlocks.ai

What is this?

Blocks.ai claims that exposing large tool sets to agents through MCP can consume roughly 26,000 tokens in tool-schema definitions before the agent processes the task, whereas CLI tools incur little or no upfront schema cost and reveal usage details on demand. The supplied snippets broadly support the underlying tradeoff: MCP pays upfront for structured tool definitions and may return tighter structured data, while CLI access preserves context at scale but can shift burdens such as output parsing, authentication, and session management onto the agent. The snippets do not independently establish Blocks.ai’s identity, methodology, or the exact 26,000-token measurement.

Why it matters to Scott

Blocks.ai independently quantifies the exact context-tax argument in Scott’s Code-First Architecture and “Why Code Execution Beats MCP,” creating a potential dated-receipts and benchmarking opportunity. The claimed 26,000-token figure could strengthen that argument, but its significance is limited until Blocks.ai’s methodology and measurement are independently established; the radar already tracks closely related MCP-schema compression work.

ip:framework.code-first-architectureip:source.why-code-execution-beats-mcpip:framework.context-engineeringradar:mcptoon-tool-discovery-compressionradar:tokencompress-agent-context-pruningradar:concept.mcp

queries asked of Scott's wikis

MCP schema overhead and context budgeting
CLI-first agent tool harnesses
on-demand tool discovery and progressive disclosure
tool-use token economics at scale
MCP versus shell tooling tradeoffs
agent handling of CLI output and authentication

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (5) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	MCP vs. CLI: 26,000 tokens burnt before your agent reads the prompt	kayleykiwi	2	1
🟧 echo.blog ⭐	Claims MCP can burn 26,000 tokens before an agent reads its prompt and compares that overhead with CLI tool access.	Blocks.ai	—	—
🟠 reddit	I built only-cli: Turn any website into a compact CLI tailored for AI agents. Browse the web in hundreds of tokens, not tens of thousands. ClaudeAI	GeekLifer	197	38
🟠 reddit	20 bucks is 20 bucks. Claude uses tokens for your connectors. ClaudeAI	iliadz	1	0
🟠 reddit	I've been dealing with the MCP side for a while, and I wanted to share what finally came up: mcpify. OpenAI	BeneficialPenalty589	1	0

Interpretation history

2026-09-03T12:29:26Z
The broader tool-schema context-tax mechanism is now established through multiple implementations, but Blocks.ai’s specific 26,000-token comparison has attracted no methodology or reproduction within its active horizon. The standalone benchmark episode has faded without changing the broader engineering conclusion.
2026-09-01T11:39:37Z
No substantive evidence has arrived: the small engagement changes neither strengthen the benchmark nor alter the established broader schema-overhead pattern. The exact 26,000-token comparison remains an unreproduced first-party claim.
2026-08-30T11:34:34Z
The MCP-side lazy-discovery implementation further establishes schema overhead as a practical engineering problem with multiple mitigation paths, not merely a CLI-versus-MCP dichotomy. It still does not independently reproduce Blocks.ai’s methodology or validate the specific 26,000-token figure.
2026-08-30T11:23:15Z
evidence attached: reddit.post.1w2es1x — The released MCP adapter's lazy tool discovery and large schema reduction provide relevant independent context on controlling tool-description context costs.
2026-08-29T03:30:38Z
Refreshed discussion is limited to implementation questions and comparisons with adjacent tools; it adds no independent reproduction of Blocks.ai’s MCP benchmark. The broader schema-overhead mechanism remains corroborated, while the specific 26,000-token claim remains unvalidated.
2026-08-28T21:40:21Z
The independent CLI implementation plus a separate report of unused connector schemas consuming context now corroborate the broader tool-schema context-tax mechanism. Neither reproduces Blocks.ai’s MCP benchmark, so the specific 26,000-token figure remains unvalidated.
2026-08-28T17:25:07Z
evidence attached: reddit.post.1w0vmxx — Reports real-world token costs from always-enabled connector schemas, independently reinforcing the case for lower-overhead tool access.
2026-08-28T12:25:32Z
The independent only-cli implementation makes the broader CLI-first context-efficiency pattern more concrete, moving the case beyond a lone conceptual claim. It does not reproduce Blocks.ai’s MCP comparison or validate the specific 26,000-token figure, so the central quantitative hypothesis remains uncorroborated.
2026-08-28T12:23:51Z
evidence attached: reddit.post.1w0osys — The released CLI artifact directly bears on whether compact CLI interfaces can reduce web-tool context and inference costs for agents.
2026-08-26T15:34:09Z
No substantive evidence has arrived beyond a trivial engagement increase; the exact 26,000-token figure remains an unvalidated first-party measurement without methodology or independent reproduction.
2026-08-26T15:30:49Z
grounded: converges/medium — Blocks.ai independently quantifies the exact context-tax argument in Scott’s Code-First Architecture and “Why Code Execution Beats MCP,” creating a potential da
2026-08-26T15:29:19Z
case created — The first-party analysis makes a specific, consequential agent-tooling cost claim, though it has little discussion or corroboration so far.

Decision trace

09-03 22:29expireThe broader tool-schema context-tax mechanism is now established through multiple implementations, but Blocks.ai’s specific 26,000-token comparison has attracted no methodology or reproduction within
09-03 22:29alert_silentThe only delta is minor engagement on previously known evidence; it adds no benchmark validation, implementation result, or decision-relevant change for Scott.
09-03 22:29alert_routeThe only delta is minor engagement on previously known evidence; it adds no benchmark validation, implementation result, or decision-relevant change for Scott.
09-01 21:39repriceNo substantive evidence has arrived: the small engagement changes neither strengthen the benchmark nor alter the established broader schema-overhead pattern. The exact 26,000-token comparison remains
09-01 21:39alert_silentThe new delta is only minor engagement and staleness, with no methodology, reproduction, or measured downstream impact; nothing would be lost by waiting for the next briefing.
09-01 21:39alert_routeThe new delta is only minor engagement and staleness, with no methodology, reproduction, or measured downstream impact; nothing would be lost by waiting for the next briefing.
08-30 21:34repriceThe MCP-side lazy-discovery implementation further establishes schema overhead as a practical engineering problem with multiple mitigation paths, not merely a CLI-versus-MCP dichotomy. It still does n
08-30 21:34alert_silentThe new implementation reinforces an already tracked context-efficiency pattern, but its schema reduction is self-reported and lacks measured downstream token, cost, or task-quality effects; it can wa
08-30 21:34alert_routeThe new implementation reinforces an already tracked context-efficiency pattern, but its schema reduction is self-reported and lacks measured downstream token, cost, or task-quality effects; it can wa
08-30 21:23alert_silentThis is a concrete open-source implementation of progressive tool disclosure and a useful additional benchmark for Scott’s context-tax argument, but the reduction is self-reported from one API example
08-30 21:23surface_candidateThis is a concrete open-source implementation of progressive tool disclosure and a useful additional benchmark for Scott’s context-tax argument, but the reduction is self-reported from one API example
08-30 21:23alert_routeThis is a concrete open-source implementation of progressive tool disclosure and a useful additional benchmark for Scott’s context-tax argument, but the reduction is self-reported from one API example
08-30 21:23attachThe released MCP adapter's lazy tool discovery and large schema reduction provide relevant independent context on controlling tool-description context costs.
08-30 21:22propose_attachThe released MCP adapter's lazy tool discovery and large schema reduction provide relevant independent context on controlling tool-description context costs.
08-29 23:21sensor_dirtyengagement_update
08-29 22:21sensor_dirtyengagement_update
08-29 17:21sensor_dirtyengagement_update
08-29 16:21sensor_dirtyengagement_update
08-29 14:21sensor_dirtyengagement_update
08-29 13:30repriceRefreshed discussion is limited to implementation questions and comparisons with adjacent tools; it adds no independent reproduction of Blocks.ai’s MCP benchmark. The broader schema-overhead mechanism
08-29 13:30alert_silentThe new comments do not add a benchmark, methodology, adoption event, or product change, so waiting for the next briefing carries no attention regret.
08-29 13:30alert_routeThe new comments do not add a benchmark, methodology, adoption event, or product change, so waiting for the next briefing carries no attention regret.
08-29 13:21sensor_dirtycomment_update
08-29 11:21sensor_dirtyengagement_update
08-29 09:21sensor_dirtyengagement_update
08-29 07:40repriceThe independent CLI implementation plus a separate report of unused connector schemas consuming context now corroborate the broader tool-schema context-tax mechanism. Neither reproduces Blocks.ai’s MC
08-29 07:40alert_silentThe new connector anecdote adds useful triangulation but no reproducible measurement, methodology, or product change; refreshed discussion around only-cli is largely amplification, so the next briefin
08-29 07:40alert_routeThe new connector anecdote adds useful triangulation but no reproducible measurement, methodology, or product change; refreshed discussion around only-cli is largely amplification, so the next briefin
08-29 07:21sensor_dirtyengagement_update
08-29 06:21sensor_dirtyengagement_update
08-29 05:21sensor_dirtyengagement_update
08-29 04:22sensor_dirtycomment_update
08-29 03:25alert_silentA single low-engagement user report adds Claude-specific anecdotal evidence that enabled account connectors and deferred tool names consume session context even when unused, but it provides no complet
08-29 03:25surface_candidateA single low-engagement user report adds Claude-specific anecdotal evidence that enabled account connectors and deferred tool names consume session context even when unused, but it provides no complet
08-29 03:25alert_routeA single low-engagement user report adds Claude-specific anecdotal evidence that enabled account connectors and deferred tool names consume session context even when unused, but it provides no complet
08-29 03:25attachReports real-world token costs from always-enabled connector schemas, independently reinforcing the case for lower-overhead tool access.
08-29 03:22propose_attachReports real-world token costs from always-enabled connector schemas, independently reinforcing the case for lower-overhead tool access.
08-29 03:21sensor_dirtycomment_update
08-29 02:22sensor_dirtycomment_update
08-29 01:22sensor_dirtycomment_update