2026-04-14

• #0006

  #0006Idea signalsHacker News · MCP Serverlast month

  `LangAlpha`: compiling `MCP` servers into Python libs to cut agent context cost

  60radar

  LangAlphaAgent framework — turns MCP into typed modules to cut tokens

  Instead of stuffing financial MCP schemas and price data into prompts, it generates typed Python modules and keeps only a one-line server summary. The stronger signal is the workspace model: persistent sandboxes plus memory/index files make long-horizon research agents much more usable.

  • —A single five-year daily price pull can dump tens of thousands of tokens; large MCP schemas alone can burn 50k+ tokens before useful work starts.
  • —The workaround is practical: auto-generate typed Python modules from MCP schemas at workspace init, then let the agent import tools like a normal library.
  • —Prompt cost stays flat even with scale: about 80 tools across servers, while only a one-line summary per server remains in context.
  • —Persistent workspaces matter more than the finance angle. Sandbox state, a memory file, and a file index survive across sessions, so analysis compounds instead of resetting.

  Source: github.com/ginlix-ai/langalphaRead original →
• #0005

  #0005Agents & toolsHacker News · Show HN AIlast month

  `LangAlpha`, a finance-focused agent harness that sidesteps MCP prompt bloat

  60radar

  LangAlphaResearch platform — turns MCP into typed Python modules

  It turns bulky MCP servers into typed Python modules and keeps research state inside persistent workspaces. Strong agent-harness ideas are reusable beyond finance; the investing layer is niche.

  • —Instead of exposing 80 tools and large schemas to the model, it generates typed Python modules at workspace init, keeping prompt cost flat as servers grow.
  • —A single request for 5 years of daily prices can dump tens of thousands of tokens; moving access into imports cuts context waste before analysis starts.
  • —Each research goal gets a persistent sandbox, plus a memory file and file index reloaded before every LLM call. Cross-session continuity is built into the runtime.
  • —Portfolio, watchlist, risk tolerance, and data-source context are injected on every call. The product layer is finance-specific, but the MCP handling pattern travels well.

  Source: github.com/ginlix-ai/langalphaRead original →
• #0004

  #0004Agents & toolsHacker News · Show HN AIlast month

  `Kontext CLI`: short-lived credential broker for AI coding agents

  60radar

  Kontext CLISecurity CLI — session-scoped creds for AI agents

  Instead of dumping API keys into .env or chat, it swaps declared placeholders for short-lived, scoped creds and streams every tool call for audit. Clean security model, but Codex support is still pending, so the immediate use case is narrower.

  • —A .env.kontext file maps services like github, stripe, and linear to placeholders, then kontext start --agent claude resolves them at runtime.
  • —For OAuth services it uses OIDC plus RFC 8693 token exchange; for static keys, the backend injects creds directly, so secrets stay memory-only during the session.
  • —Each tool call is logged with action, result, allow status, user, session, and org attribution, which fixes the usual blind spot after handing raw keys to an agent.
  • —The CLI is written in Go, adds about 5ms hook overhead per tool call, talks to the backend over ConnectRPC, and stores auth in the system keyring.
  • —Server-side allow/deny enforcement is not closed yet. Audit exists now; hard policy blocking is the next step.

  Source: github.com/kontext-dev/kontext-cliRead original →
• #0003

  #0003Idea signalsHacker News · AI Agentlast month

  `LangAlpha`: domain-specific agent harness for persistent research

  50radar

  LangAlphaAgent platform — vertical harness for persistent research

  The reusable part is not finance UI but the agent architecture: compile MCP schemas into importable Python modules and keep research state in a persistent workspace. Useful signal if you're building vertical agents where context cost and cross-session memory break generic chat UX.

  • —MCP schema bloat is the first bottleneck: tool definitions alone can cost 50k+ tokens, so they generate typed Python modules at init and keep only a one-line server summary in prompt.
  • —Tool count stops mattering once the agent imports local wrappers like a normal library; they claim roughly 80 tools with similar prompt cost whether a server exposes 3 or 30 tools.
  • —Each research goal gets its own persistent sandbox, memory file, and file index, then that state is reloaded before every LLM call. This fixes the usual 'start over next session' failure mode.
  • —The stronger wedge is vertical context injection: portfolio, watchlist, risk tolerance, and data sources travel with every call. Same pattern fits SaaS ops, growth, or support agents.

  Source: github.com/ginlix-ai/langalphaRead original →
• #0002

  #0002Agents & toolsHacker News · MCP Serverlast month

  `Kontext CLI`: Short-lived credential broker for coding agents

  50radar

  Kontext CLICLI tool — injects session creds via OIDC

  Instead of pasting long-lived secrets into .env, it swaps placeholders for session-scoped creds and streams every tool call for audit. Strong pattern for agent-era auth, but immediate value depends on whether your stack fits its supported integrations.

  • —Projects declare needs in .env.kontext, then kontext start --agent claude resolves GitHub, Stripe, Linear creds at runtime with no disk write.
  • —OAuth services use RFC 8693 token exchange; static keys are injected only into process memory. That removes raw secret sprawl but centralizes trust in the backend.
  • —Each tool call is attributed to user, session, and org, with allow/deny plumbing already wired. Useful if agent actions need audit trails before broader team use.
  • —Built in Go with roughly 5ms hook overhead per tool call, ConnectRPC backend links, and system keyring auth storage. Performance cost looks low enough for daily CLI use.

  Source: github.com/kontext-dev/kontext-cliRead original →
• #0001

  #0001Agents & toolsHacker News · LLMlast month

  Show HN: `CodeBurn` analyzes `Claude Code` token usage by task

  50radar

  CodeBurnAnalytics tool — categorizes Claude Code token use by task

  CodeBurn is a local analysis tool for Claude Code session transcripts that groups usage into 13 task categories without extra LLM calls. For indie developers, its main value is exposing where spend actually goes, including the claim that 56% of cost came from non-tool conversation and only 21% from code edits/writes.

  • —Built to solve poor cost visibility in Claude Code, after reported spending of ~$1400/week.
  • —Reads local JSONL transcripts from ~/.claude/projects/ and classifies turns by tool-usage pattern with no additional LLM cost.
  • —Indie-dev relevance: helps identify prompt/process waste, not just daily or per-model totals like ccusage.
  • —Product differentiation comes from an interactive TUI built with Ink, plus SwiftBar integration for quick monitoring on macOS.

  Source: github.com/AgentSeal/codeburnRead original →