Let your agent monitor and optimize its own LLM cost.Beta

Three parts. One account. Point your OTel traces at us, open a dashboard, and connect the agent itself to the same data via MCP + skill — so it can propose, benchmark and apply its own cost reductions.

01 · Ingest

Point your existing OpenTelemetry pipeline at us. Works with Claude Code (the CLI), LangChain, LlamaIndex, CrewAI, Mastra, Vercel AI SDK, Vertex AI, Bedrock, or any custom code that uses an OTel-instrumented OpenAI / Anthropic / Google SDK.

Claude Code (CLI) — env vars only

Claude Code v2.1+ emits OpenTelemetry metrics for every LLM call out of the box. Export these env vars (easiest: put them in a claude shell wrapper or a per-project .envrc) and your spend starts flowing in on the next session:

Copyexport CLAUDE_CODE_ENABLE_TELEMETRY=1
export OTEL_EXPORTER_OTLP_ENDPOINT=https://adapterly.ai/ingest
export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
export OTEL_METRICS_EXPORTER=otlp
export OTEL_LOGS_EXPORTER=otlp
export OTEL_EXPORTER_OTLP_HEADERS="X-API-Key=adk_your_key_here"

No forwarder, no proxy, no code changes. We parse Claude Code's native claude_code.token.usage + claude_code.cost.usage metrics and fold them into the same dashboard as all other sources. Backfill of past sessions is optional — see the forwarder script.

OTel-instrumented app code

For code that already uses the OTel SDK (LangChain, OpenAI SDK + openinference, custom instrumentation), point traces at us:

CopyOTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://adapterly.ai/ingest/v1/traces
OTEL_EXPORTER_OTLP_TRACES_HEADERS=X-API-Key=adk_your_key_here

We accept both OTLP protocols — http/protobuf (the OTel SDK default) and http/json. No protocol override needed.

If you don't yet have an OTel SDK wired up, one package + three lines gets you going:

Copy# Python example — same pattern for OpenAI / Anthropic / LangChain
pip install opentelemetry-sdk opentelemetry-exporter-otlp-proto-http \
            openinference-instrumentation-openai

# in your app startup:
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from openinference.instrumentation.openai import OpenAIInstrumentor

trace.set_tracer_provider(TracerProvider())
trace.get_tracer_provider().add_span_processor(
    BatchSpanProcessor(OTLPSpanExporter())  # reads OTEL_* env vars
)
OpenAIInstrumentor().instrument()
# ...your existing OpenAI client code, unchanged...

02 · Dashboard

Spend-per-day chart, top models / projects / users, cache-read ratio, and rule-based recommendations. API key stays in your browser (localStorage) — nothing is sent anywhere except to this dashboard's backend.

→ Open the dashboard

03 · MCP + Skill — the agent self-optimizes

This is the part that makes Adapterly different. The same data your dashboard shows, the agent can read itself through four MCP tools:

• my_spend_today() — quick status check
• my_spend_window(days=N) — trend + monthly projection
• my_breakdown(by='model'|'project'|'user') — find the cost drivers
• my_recommendations() — concrete reduction ideas with impact estimates

Connect from Claude Code

Copyclaude mcp add --transport http adapterly-ingest \
  https://adapterly.ai/ingest/mcp/ \
  --header "X-API-Key: adk_your_key_here"

Connect from Cursor / VS Code / any MCP client

Copy{
  "mcpServers": {
    "adapterly-ingest": {
      "url": "https://adapterly.ai/ingest/mcp/",
      "headers": {"X-API-Key": "adk_your_key_here"}
    }
  }
}

Install the agent skill

The skill tells the agent when to call which tool and how to interpret the result. For Claude Code, drop it under ~/.claude/skills/:

Copymkdir -p ~/.claude/skills
curl -o ~/.claude/skills/adapterly-cost-aware.md \
     https://adapterly.ai/ingest/skill.md

Other clients: paste the skill markdown as a system prompt or custom instruction.

What it looks like in chat

The agent treated the recommendation as a hypothesis, pulled context from its own session, and asked you before changing anything — following the skill guide. That's the loop.

Why this works

• Vendor-neutral — Claude, GPT, Gemini, Groq, Together, Fireworks all in one view
• Out-of-band — no proxy latency, no code changes to your LLM calls
• EU-hosted — backend in Hetzner Falkenstein, GDPR-friendly by default
• Agent-first — the only platform that gives the agent itself read access to its cost data via standard MCP
• Metadata-only — we store model, token counts, user id — never prompt bodies or responses

Pricing

Free during our beta. Terms: honest feedback and, if the data is useful, permission for an anonymized case study.

Post-beta plans (indicative):

• Free — up to 50k spans / month, 30 days retention. Individual developers.
• Team — $15–30 / user / month. Multi-user auth, Slack alerts, 90 days retention.
• Enterprise — custom. SSO, RBAC, SOC 2 posture, on-prem option, SLA.