BurnLens — The open-source AI economics control plane

Measure, explain and control the cost of AI agents and LLM applications. Connect spend to repositories, teams, customers, features and outcomes across OpenAI, Anthropic, Google, Groq, Mistral, Together, xAI, DeepSeek, Azure OpenAI, and AWS Bedrock.

BurnLens is the economic control plane for AI: Measure. Control. Prove. Savings marked observed are reductions in cost per request during a measured follow-up window; outcome quality and invoice-level savings require additional evidence.

pip install burnlens
burnlens scan
# cost by repository, then optional: burnlens cloud connect

The Problem

Bills tell you the model, not the why. Your invoice says gpt-4o: $4,287. It doesn't say which feature, which team, or which customer burned it. By the time you trace the spike, it's already on next month's card.

Alerts arrive after the damage. A bad deploy, a runaway agent, or one abusive customer can trigger thousands of API calls before any dashboard turns red. You find out when you open the bill — or when your CEO does.

Every provider is a different silo. OpenAI's usage page. Anthropic's console. Azure Cost Management. Bedrock CloudWatch. No unified view, no way to ask "which feature is our biggest AI spend across all providers."

How It Works

1. Drop-in proxy. Point your SDK's BASE_URL at localhost:8420. Existing code works unchanged. The proxy is designed for low overhead and supports streaming passthrough.

2. Tag what matters. Request headers (X-BurnLens-Tag-Feature, X-BurnLens-Tag-Team, X-BurnLens-Tag-Customer, plus X-BurnLens-Tag-Agent-Id and X-BurnLens-Tag-Workflow-Id for agent workloads) attribute any call to any dimension. Tags are stripped before the request reaches the AI provider. If you enable cloud sync, tag values are uploaded to your workspace alongside cost metadata — never prompt or response bodies.

3. Control before you call. Register an API key with a daily dollar limit. At 100% of recorded spend, BurnLens returns 429 before the upstream request is made. 50% and 80% thresholds fire Slack or email alerts. Run burnlens controls to inspect scope and concurrency guarantees; exact behaviour is specified in Budget enforcement semantics.

4. One dashboard for supported providers. OpenAI, Anthropic, Google, Groq, Mistral, Together, xAI, DeepSeek, Azure OpenAI, and AWS Bedrock spend in one unified view. Model breakdowns, waste detection, and budget tracking use versioned provider pricing.

Code Example

import os, openai

os.environ["OPENAI_BASE_URL"] = "http://127.0.0.1:8420/proxy/openai"

client = openai.OpenAI(default_headers={
    "X-BurnLens-Tag-Feature": "chat",
    "X-BurnLens-Tag-Team": "backend",
    "X-BurnLens-Tag-Customer": "acme-corp",
})

Tags are stripped before the request reaches OpenAI. They never appear in the provider API payload.

Agent workloads. Tag with X-BurnLens-Tag-Agent-Id and X-BurnLens-Tag-Workflow-Id to get cost per agent and cost per workflow instead of only cost per model:

client = openai.OpenAI(default_headers={
    "X-BurnLens-Tag-Agent-Id": "refund-agent",
    "X-BurnLens-Tag-Workflow-Id": "refund_review",
})

Break spend down with GET /api/costs/by-tag?tag=agent_id locally, or GET /api/v1/usage/by-tag?tag_type=agent_id on BurnLens Cloud. Tool and function calls are counted per request, so a looping agent shows up as tool-call volume rather than just a larger bill.

Multi-word tag headers accept either spelling — X-BurnLens-Tag-Agent-Id or X-BurnLens-Tag-Agent_Id. Prefer the hyphenated form: nginx drops headers containing underscores unless underscores_in_headers is on.

Cost per accepted outcome

Spend alone doesn't tell you whether it was worth it. Report what a workflow produced and BurnLens divides one by the other:

burnlens outcome record --workflow refund_review --status accepted --id ticket-8412
burnlens outcome show

 Workflow        Accepted  Rejected/Failed  Total cost  Rework   Unattributed  Per accepted
 refund_review   780       140              $18.40      $3.20    $1.10         $0.0236

From an application, post outcomes to BurnLens Cloud instead:

curl -X POST https://api.burnlens.app/v1/outcomes \
  -H "X-API-Key: $BURNLENS_API_KEY" \
  -d '{"outcomes":[{"outcome_id":"ticket-8412","workflow_id":"refund_review",
                    "status":"accepted","event_time":"2026-08-09T12:00:00Z"}]}'

Read the economics back with the same key — no browser session needed:

curl https://api.burnlens.app/api/v1/outcomes/summary \
  -H "X-API-Key: $BURNLENS_API_KEY"

Three things worth knowing about the number:

• outcome_id is yours and it is the idempotency key. Re-posting the same one is ignored, so at-least-once delivery and importer re-runs can't inflate the count that the cost is divided by.
• Per accepted = total workflow spend / accepted outcomes. Failed and rejected attempts are charged to the successes, because that is what one working result actually costs. The Rework column shows how much of it was spent on attempts that didn't land.
• Unattributed spend is shown, not hidden. A request is charged to the first outcome of its workflow that follows it within a window (24h by default, --window to change). Spend with no outcome after it stays visible in its own column rather than quietly disappearing from the denominator.

If an outcome genuinely changes, use burnlens outcome correct <outcome-id> --status accepted --reason "..." (or Cloud PATCH /v1/outcomes/{outcome_id}); the current result stays one count and the prior status remains available through burnlens outcome history <outcome-id>.

A workflow with spend and no accepted outcomes reports no unit cost at all rather than $0 — the absence is the signal.

Coding agents: no instrumentation required

For agent work you don't have to report anything. A merged pull request already is an accepted outcome, and a closed-unmerged one is a rejected outcome — so BurnLens reads them out of GitHub and joins them to the agent spend it scanned off disk:

burnlens scan --provider claude     # agent session cost, then merged PRs if `gh` is installed
burnlens outcome show               # cost per merged PR

burnlens scan imports local session logs, then derives outcomes for the current checkout through the GitHub CLI. If gh is not on PATH, scan still imports cost and prints that fact — it does not skip the step silently. burnlens outcome derive remains the command for a later re-run or a different checkout.

Measured on this repository while building it, over the 2026-07-10 to 2026-08-15 telemetry window, on 2026-08-15: 104 merged PRs, 2 closed unmerged, $710.85 of agent spend — about $6.84 per accepted PR. Failed attempts are charged to the successes, which is what one merged PR really costs. (A floor, not a ceiling — any model missing from the pricing tables is imported as $ unknown, so check burnlens pricing if a number looks low.)

Import and derive are both idempotent and safe on a schedule: outcome ids are derived deterministically from the repo and PR number, so re-running only ever adds newly-closed PRs.

Cost is attributed per repository rather than per PR, because agent session logs record which repo a session ran in, not which branch. With several PRs in flight that is the honest reading of what one merged PR costs.

Use Cases

Coding agents. Cursor, Claude Code, Codex, Gemini CLI — cost per repo and developer from local logs. Merged PRs become accepted outcomes when gh is present (repo-level, because session logs do not record the branch). Per-PR tags apply to live proxy traffic via burnlens run. Set a hard daily cap per API key so one runaway agent can't blow the team's monthly budget overnight.

Customer-facing AI. Tag each request with a customer ID. See which customers drive the most cost, alert on thresholds, and optionally route to cheaper models.

RAG and agents. Tag retrieval calls, tool calls, and generation separately. See whether your vector search or synthesis step is the cost driver — and whether it justifies the output quality.

Internal tools. Set per-team monthly budgets, get Slack alerts at 80% and 100%, and export monthly records for comparison with provider invoices.

Supported Providers

Pricing covers current text/chat models for the supported providers, plus audio-modality tokens (OpenAI *-audio-preview / *-realtime-preview, billed at their own per-million rate) and arbitrary flat per-unit fees via each model's optional unit_prices (e.g. per web-search call). Image and video generation are still out of scope. Gemini Pro's higher rate above 200K input tokens is applied automatically (tiered pricing, since v1.12.0). Audio rates should be re-checked against the provider pricing page — they change less often than text but do move.

Why BurnLens

Dashboard

Configuration

Zero config required — sensible defaults out of the box. Optional burnlens.yaml:

budget_limit_usd: 500.00
budgets:
  teams:
    backend: 200.00
    research: 100.00
  customers:
    acme-corp: 50.00
alerts:
  slack_webhook: https://hooks.slack.com/...
  # Per-agent baselines: alert when an agent tagged `agent_id` burns 3x its own
  # 7-day hourly average, or loops on one trace_id. Defaults shown.
  agent_deviation_multiplier: 3.0
  agent_min_spend_usd: 1.00
  agent_loop_max_requests: 20
  agent_loop_window_minutes: 10

CLI

burnlens start                  # proxy + dashboard on :8420
burnlens top                    # live cost by model (htop-style)
burnlens report                 # weekly cost summary
burnlens analyze                # waste detection report
burnlens export                 # CSV of last 7 days
burnlens run -- python app.py   # auto-tag a process with repo / dev / pr / branch
burnlens key register <name>    # label an API key + set a daily cap
burnlens key list               # list registered keys with caps
burnlens keys                   # today's spend per registered key
burnlens scan --provider claude # import Claude Code session costs from disk
burnlens scan --provider cursor # import Cursor IDE session costs from disk
burnlens scan --provider codex  # import OpenAI Codex session costs from disk
burnlens scan --provider gemini # import Gemini CLI session costs from disk
burnlens repos                  # cost by repository
burnlens cloud connect          # write cloud config (never prompts)
burnlens sync --now             # push unsynced cost metadata

Contributing

Issues and PRs welcome. See CONTRIBUTING.md.

git clone https://github.com/sairintechnologycom/burnlens
cd burnlens
pip install -e ".[dev]"
pytest

License

Apache License 2.0