SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Wie viele Tokens die MCP-Tool-Liste von osFoundry in Codex kostet und wie man sie pro Projekt kürzt

0
·2 views

Ein Entwickler hat den Tool-Katalog des osFoundry-MCP-Servers per curl abgefragt und dabei 59 Tools mit einer Antwort von rund 45 KB ermittelt. Codex CLI bindet den Server über eine Konfigurationstabelle ein, und die Tool-Deklarationen werden danach bei jedem Modell-Request mitgeschickt. Gemessen mit dem Tokenizer tiktoken (Encoding o200k_base) belegt der Payload etwa 7.600 Tokens, wobei Beschreibungen und Input-Schemas den größten Anteil ausmachen. Der Server kündigt Änderungen an der Liste während einer laufenden Session nicht an, sodass die beim Start übermittelte Liste fix bleibt. Der Artikel erläutert daher, wie sich die Liste pro Projekt eingrenzen lässt, um den Token-Verbrauch zu senken.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Avoiding PostgreSQL Row-Level Lock Deadlocks Through Ordered Acquisition and NOWAIT

High-throughput systems that update interdependent rows concurrently often hit PostgreSQL deadlocks (SQLSTATE 40P01) or lock timeouts (SQLSTATE 55P03). The article argues that when workers acquire FOR UPDATE locks in inconsistent orders, circular waits form and the database aborts one transaction. Its proposed remedy is to sort lock keys deterministically before querying, which removes circular dependencies. Pairing this with NOWAIT and structured exception handling lets workers fail fast and roll back rather than exhaust connection pools. The piece also includes a Python example that attempts to verify this behavior across threads, along with a runtime dependency installer that the author presents as a convenience.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Meta ends new hardware features for Quest 2 and Pro, but software updates keep improving vision training apps

Meta has announced it will stop adding new hardware features to the Quest 2 and Quest Pro headsets. A developer building a vision training app reports that recent Quest OS updates have still improved performance, including better wrist tracking accuracy and lower jitter. By adjusting the polling interval in Unity's XR Interaction Toolkit, the developer cut gaze-guidance latency from 30 ms to 12 ms. A newer Scene Understanding API let them add real-time obstacle avoidance in about a week without native plugins, also shrinking build size by 3 MB. The developer plans to benchmark each OS update and automate regression testing in CI.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

RepoXray: Browser-Based Scanner Flags Supply-Chain and AI Agent Security Risks

A developer has released RepoXray, a tool for auditing code repositories before installing packages or attaching AI coding agents. It targets 'slopsquatting', where attackers register package names that large language models commonly hallucinate, by checking npm registry existence, package age under 21 days, and OSV.dev advisories. The scanner also flags risky postinstall scripts, plaintext tokens in MCP configuration files, and .env files that coding agents could read. Its web version runs entirely in the browser with no uploads, while the 19 KB CLI, run via npx rxray, has no runtime dependencies. The author is seeking technical feedback on false positives and missing detection rules.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

DeeperAI: a Mac, iPhone and iPad app that organizes AI chats into topic trees

DeeperAI is a Mac, iPhone and iPad app that lets users structure a subject as a tree of themes and sub-themes, with Claude available inside each node. Unlike a single long chat history, each conversation stays tied to its topic, so users can resume where they left off. Themes can hold notes and attachments, which are saved in folders mirroring the tree and stay in sync with Finder. The app's AI can suggest sub-themes and emojis, and users can save answers as attachments or import them as subtopics. This first part of a series explains the app's features, while setup is covered in a later installment.

0 comments Read more at DEV Community