Cost & Budget Management

AI inference costs can spiral fast when agents run unsupervised. Open Astra solves this with real-time cost dashboards, per-agent budget caps, and automatic enforcement — so you always know exactly what you're spending and no single agent can blow your budget. Most teams reduce their inference costs 30–50% in the first month just by seeing the breakdown.

Why this matters

• Predictable spend — hard caps prevent runaway agents from burning through your inference budget overnight
• Per-agent attribution — know exactly which agent is costing you money, down to the tool call
• Provider visibility — see spend by provider (Claude, OpenAI, Groq, etc.) to optimize your provider mix
• Budget inheritance — when agents spawn sub-agents, budgets flow downward with strict limits

Cost dashboard

The cost dashboard aggregates spend across your workspace by agent, provider, and time period. The default lookback is 3 months.

# Get workspace cost summary (default: last 3 months)
curl http://localhost:3000/costs \
  -H "Authorization: Bearer ${JWT_TOKEN}"

# Response
{
  "workspaceId": "ws_abc123",
  "totalCost": 42.17,
  "byAgent": {
    "research-agent": 18.50,
    "code-agent": 14.22,
    "support-agent": 9.45
  },
  "byProvider": {
    "claude": 28.90,
    "openai": 10.12,
    "groq": 3.15
  },
  "byMonth": [
    { "month": "2026-01", "cost": 15.30 },
    { "month": "2026-02", "cost": 14.80 },
    { "month": "2026-03", "cost": 12.07 }
  ]
}

Detailed breakdown

For finer-grained analysis, use the breakdown endpoint with optional date range filtering.

# Detailed breakdown with date range
curl "http://localhost:3000/costs/breakdown?startDate=2026-02-01&endDate=2026-02-28" \
  -H "Authorization: Bearer ${JWT_TOKEN}"

# Response
{
  "meta": {
    "workspaceId": "ws_abc123",
    "months": 3,
    "startDate": "2026-02-01",
    "endDate": "2026-02-28",
    "generatedAt": "2026-03-07T12:00:00.000Z"
  },
  "totalCost": 14.80,
  "byAgent": { ... },
  "byProvider": { ... },
  "byTool": { ... }
}

Budget caps

Every agent can have resource constraints defined in astra.yml or via the API. Budgets control tokens, cost, tool calls, execution time, and spawn permissions.

# astra.yml — per-agent budget constraints
agents:
  - id: research-agent
    budget:
      maxTotalTokens: 50000
      maxCostCents: 50        # $0.50 per request
      maxToolCalls: 20
      maxDuration: 60000      # 60 seconds
      maxChildAgents: 0       # cannot spawn sub-agents

System hard caps

These limits cannot be exceeded by any agent or swarm, regardless of configuration:

# System hard caps (cannot be exceeded by any agent)
maxPromptTokens:     500,000
maxCompletionTokens: 100,000
maxTotalTokens:      600,000
maxCostCents:        500        # $5.00 per swarm
maxToolCalls:        100
maxDuration:         300,000 ms  # 5 minutes
maxSpawnDepth:       5
maxChildAgents:      20

Role-based defaults

When an agent has a role but no explicit budget, these defaults apply:

# Role-based defaults (applied when no explicit budget is set)
researcher:  50,000 tokens  |  $0.50  |  20 tool calls  |  60s   |  no spawn
analyst:     80,000 tokens  |  $1.00  |  15 tool calls  |  90s   |  no spawn
writer:      30,000 tokens  |  $0.30  |   5 tool calls  |  30s   |  no spawn
mediator:    40,000 tokens  |  $0.40  |   5 tool calls  |  60s   |  no spawn

Budget inheritance

When a parent agent spawns a sub-agent, budgets follow strict inheritance rules:

1. Child budget is always ≤ parent budget (hard constraint)
2. maxSpawnDepth decrements by 1 at each level
3. Tool permissions: child gets the intersection of parent's allowed tools and union of denied tools
4. Memory permissions use AND logic — child can only have what parent allows

Cost ledger

The cost ledger tracks execution metrics at two levels:

Leaderboard

The leaderboard ranks agents by usage and performance. Useful for identifying which agents consume the most resources and which perform best.

Differential privacy budget

For workspaces with differential privacy enabled, Open Astra tracks a privacy budget that limits the amount of information that can be extracted about any individual user. Owners can reset the budget when needed.

# Get differential privacy budget status
curl http://localhost:3000/security/dp-budget \
  -H "Authorization: Bearer ${JWT_TOKEN}"

# Reset DP budget (owner only)
curl -X POST http://localhost:3000/security/dp-budget/reset \
  -H "Authorization: Bearer ${JWT_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{ "uid": "uid_alice" }'

Related

• Budget Pre-Flight — how budgets are checked before inference
• Quotas — per-agent rate limits and token quotas
• Cost Tagging — per-tool cost attribution