Costzera docs

Point Costzera at your agents' traces and code. It continuously analyzes every run for cost inefficiencies, writes recommendations grounded in your actual code, and — once you apply a fix — tracks the change against a baseline to prove the delta.

The loop is Connect → Analyze → Implement → Measure, and it runs on autopilot or with a human in the approval seat.

LangSmith

In onboarding, paste a LangSmith Personal Access Token to auto-discover your workspaces. For each workspace you want synced, add a scoped service key (or fall back to the token). Production workspaces ingest daily; test/staging workspaces are registered but stay off until you enable them.

GitHub

Install the Costzera GitHub App (read-only) on the repositories your agents live in. This lets Costzera ground its recommendations in your real code and attribute cost changes to commits. See Savings & measurement for how commits map to measured savings.

Costzera classifies inefficiency into three layers:

Each recommendation carries the evidence pulled from your traces (model, token profile, duplicate-input rate, cost share…), a visible savings calculation, and a confidence score. Three ways to act on one:

• Copy prompt — hands a ready-to-paste implementation prompt to your coding agent (Claude Code, Copilot, ChatGPT), assembled from the fix, the relevant code, and the evidence.
• Generate fix — produces a before/after code diff you can review.
• Approve / Reject / Mark implemented — moves the recommendation through its lifecycle.

When you ship a Costzera fix, record it in a costzera.json manifest at your repo root, committed in the same change. Each entry ties a recommendation to the commit that implemented it:

{
  "version": 1,
  "applied": [
    {
      "recommendation_id": "b1f2…-…-…",
      "agent": "billing_assistant",
      "pattern": "tool_pruning",
      "commit": "abc1234",
      "applied_at": "2026-07-16T02:14:00Z"
    }
  ]
}

Costzera then measures the change: it takes the agent's cost-per-run in the window before the commit as the baseline, watches the next 7 days, normalizes for traffic volume, and records the realized monthly saving — but only if quality (success rate) holds. That measured number is what shows up as verified savings.

Costzera is built on a simple principle: we analyze your agent's behaviour without taking custody of your data. We never store your prompts, completions, tool payloads, or source code.

Use Costzera from any MCP client — Claude, Claude Code, ChatGPT connectors, GitHub Copilot — right alongside the dashboard.

• Endpoint — POST https://costzera.com/api/mcp
• Auth — create a key in Settings → API keys, then send Authorization: Bearer czk_…

Tools

Read (any key): get_cost_summary, list_agents, list_recommendations, get_recommendation, get_verified_savings, get_sync_status, list_analysis_runs, get_analysis_run.

Write (read+write key only): run_analysis, approve_recommendation, reject_recommendation, implement_recommendation, set_agent_budget.

Prompts (ready-made workflows): cost_review, triage_recommendations, investigate_spike.

Connect

Claude Code:

claude mcp add --transport http costzera \
  https://costzera.com/api/mcp \
  --header "Authorization: Bearer czk_your_key"

GitHub Copilot — .vscode/mcp.json:

{
  "servers": {
    "costzera": {
      "type": "http",
      "url": "https://costzera.com/api/mcp",
      "headers": { "Authorization": "Bearer czk_your_key" }
    }
  }
}

ChatGPT — add an MCP connector with the same URL and bearer header.

The Costzera Skill teaches your agent to drive these tools into higher-level flows — a weekly cost review, triaging open recommendations, investigating a spend spike, or implementing a fix — with guardrails: never invent numbers, always confirm before a write action.

Drop the skills/costzera folder into .claude/skills/ (Claude Code) or upload it as a Skill in the Claude apps.