SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Respan Launches Open-Source Rust AI Gateway Built on a Plugin Architecture

0
·4 views

Respan, a Y Combinator W24 startup, has launched an open-source AI gateway written in Rust that uses a small core with everything else built as plugins. The design covers LLM providers, protocols such as MCP and voice, modalities like images and video, and features including fallbacks, PII redaction, and smart routing. The company says it grew out of a wrapper around LiteLLM that became hard to maintain, and that a failed LiteLLM upgrade in January led it to redesign the gateway. Developers can add their own providers and features as plugins without changing the core, and the gateway translates between API formats so applications can use models from other vendors. The project is inspired by Nginx's minimal-core approach.

Read the full story at Hacker News

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 ·

Claude Code 2.1.304 Fixes Worktree Bug That Could Delete Branches on Exit

Claude Code's --worktree option normally asks before removing a worktree that contains work, but a bug could remove it and its branch without prompting. The release note says this happened when HEAD was detached or pointed at a branch other than the one Claude created, which could lose commits. Version 2.1.304, published by Anthropic on October 19, fixes the problem. The article's author notes the affected versions and root cause were not specified, and they did not reproduce the bug. Users who used --worktree are advised to update and check for recently lost commits using git fsck, then save any found commits to a branch.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AI Agent Registers Domain and Connects It to Sentinel App on Vercel for $1.54

A software factory asked an AI agent to handle a small infrastructure task for its Sentinel app, which was already deployed on Vercel. The agent registered sentinelproof.click after the preferred sentinel.click name turned out to be taken, paying $1.54 for a one-year term from a $10 prepaid balance. Before buying, it ran a dry-run registration check, then redeployed Sentinel and confirmed its health endpoints were working. The domain was attached to the Vercel project, and after DNS propagation and certificate issuance, Sentinel loaded over HTTPS. The team also built in safeguards such as a spending ceiling, replay limits, and encrypted operation state, while noting that a separate safety-report feature was limited by a paid upstream service.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Re-running stock asset batches can get creators flagged, and a tracking script prevents it

A contributor describes uploading a batch of 60 sticker assets, then accidentally re-running the same upload command, which produced 120 duplicate submissions and a review hold on their Adobe Stock account. Stock platforms reportedly fingerprint files for visual similarity and treat repeated submissions as a behavior problem, with pattern duplicates leading to longer reviews and quality-based rejections. To address this, the author built a simple text-ledger script that records completed uploads and lets interrupted batches resume without reprocessing finished files. A companion script also generates the exact CSV metadata headers that Adobe Stock, Vecteezy, and Dreamstime each require. The tools are built with standard libraries to run on modest hardware.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Five agent job boards tested: how payouts are made and where they can quietly fail

An autonomous agent published its measured results from working on several agent job boards, noting it had received no payments yet. It describes Taskmarket, which escrows USDC on Base and settles awards as on-chain transfers, with a public example payment. Arena42 is presented as the cleanest option, but unbound wallets can cause rewards to be skipped, and WURK's token balance is not spendable money. MusedIn shows far fewer paid jobs than hires, while BountyBook's scoring failed despite passing local checks. The author also highlights Frantic's transparent API, claim rules and delivery preflight as practices worth copying.

0 comments Read more at DEV Community