Releases·Tool Release
Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI
Cast AI presented Kimchi, its open-source coding harness, at AI Engineer World's Fair 2026. Instead of rationing tokens, the harness measures cost per task and automatically selects proprietary or open models per task based on outcomes. It also ships Ferment for multi-hour autonomous coding runs, Teleport for persistent agent sessions, and Studio for team collaboration.
22 engagement1 sourceFri, Oct 2, 2026, 08:00 PM
AI Overview
- Cast AI built open-source coding harness Kimchi after coding-agent costs soared, presenting at AI Engineer World's Fair 2026 in San Francisco
- Co-founder Laurent Gil and engineering lead Žilvinas Urbonas advocate measuring cost per task instead of cost per token
- The outcome-aware harness picks proprietary or open models per task and switches automatically as new models ship
- Internal use by 300 employees over three months delivered 2.5x savings
- Ferment runs milestone-based, self-scoring multi-hour autonomous coding tasks that deploy to staging
- Teleport keeps agent sessions running remotely after closing a laptop; Studio adds a shared Kanban board for team collaboration
Related
Products·launched
Tokenless – API gateway for automatic model switching to cut AI costs
Tokenless is an API gateway that dynamically routes agent traffic between different LLMs on a turn-by-turn basis to reduce AI spend. It helps companies like Uber and Salesforce avoid blowing their yearly AI budget by using cheaper open-source models when they suffice, reserving expensive frontier models only when needed.
Tokenless is an API gateway that dynamically routes agent traffic between different LLMs on a turn-by-turn basis to reduce AI spend. It helps companies like Uber and Salesforce avoid blowing their yearly AI budget by using cheaper open-source models when they suffice, reserving expensive frontier models only when needed.
126 engagement1 sourcehackernewsWed, Jul 29, 2026, 03:55 PM
Products·launched
TrueForge: open-source model-neutral agent harness matching Claude Managed Agents accuracy at lower cost
TrueForge is an open-source, model-neutral harness for running general-purpose agents with any LLM. The team benchmarked it against Claude Managed Agents on DevRev Enterprise-Bench, matching 11/14 solve rates with the same Opus 4.8 model while using ~63% fewer tokens and lower cost per run. It targets developers who want harness-level control without vendor lock-in.
TrueForge is an open-source, model-neutral harness for running general-purpose agents with any LLM. The team benchmarked it against Claude Managed Agents on DevRev Enterprise-Bench, matching 11/14 solve rates with the same Opus 4.8 model while using ~63% fewer tokens and lower cost per run. It targets developers who want harness-level control without vendor lock-in.
92 engagement1 sourceredditThu, Sep 3, 2026, 12:07 PM
Products·building
Felan: open-source, model-portable coding agent that reduces model spend
Felan is an open-source coding agent built on Pi that completes software work correctly while spending less on models. It uses model routing, progressive context, compact tool output, and explicit task state to cut wasted tokens. It targets developers who want lower cost without accepting cheaper-but-failed runs.
Felan is an open-source coding agent built on Pi that completes software work correctly while spending less on models. It uses model routing, progressive context, compact tool output, and explicit task state to cut wasted tokens. It targets developers who want lower cost without accepting cheaper-but-failed runs.
15 engagement1 sourceredditTue, Sep 8, 2026, 06:09 AM
Releases·Tool Release
Developer builds open-source 'manager' pattern for AI coding agents, cutting costs 35-41%
A solo developer implemented a pattern from Cognition (the Devin team) where an expensive frontier model acts as a 'manager' that plans and reviews code changes but cannot edit files directly, delegating all mechanical work to a cheaper, faster model. The approach, published as open source for the opencode terminal coding agent, reportedly reduces costs by 35-41% while maintaining quality. Subsequent posts show the pattern extended to other tools like Claude Code, with users reporting token reductions of 58.6% and cost cuts of 45.8% on SWE-bench Lite using Codex 5.6 as manager and cheaper models as executors.
A solo developer implemented a pattern from Cognition (the Devin team) where an expensive frontier model acts as a 'manager' that plans and reviews code changes but cannot edit files directly, delegating all mechanical work to a cheaper, faster model. The approach, published as open source for the opencode terminal coding agent, reportedly reduces costs by 35-41% while maintaining quality. Subsequent posts show the pattern extended to other tools like Claude Code, with users reporting token reductions of 58.6% and cost cuts of 45.8% on SWE-bench Lite using Codex 5.6 as manager and cheaper models as executors.
Updated Mon, Jul 20, 2026, 12:19 AM — Users report extending the pattern to Claude Code and Codex 5.6, achieving up to 58.6% token reduction and 45.8% cost sa
30089 engagement3 sourcesredditSun, Jul 12, 2026, 05:35 PM
Products·building
Hungry: coding harness for free AI models with auto-switching
Hungry is a coding tool that optimizes the use of free AI models by providing a harness with features like OpenAI-compatible API integration, model pools that automatically switch when limits are hit, and token-saving. It solves the problem of managing multiple free AI model quotas and improving efficiency for developers who rely on free tiers.
Hungry is a coding tool that optimizes the use of free AI models by providing a harness with features like OpenAI-compatible API integration, model pools that automatically switch when limits are hit, and token-saving. It solves the problem of managing multiple free AI model quotas and improving efficiency for developers who rely on free tiers.
1 engagement1 sourceredditMon, Aug 17, 2026, 11:44 AM