Llmkit

LLMKit is an open-source AI gateway and SDK suite for cost attribution, budget admission, and request evidence. The gateway reserves estimated spend before provider dispatch. It rejects requests that cannot fit the active budget, then settles admitted reservations to actual usage when the response completes.

The repository also ships local tracking surfaces that do not require an LLMKit account or proxy.

Choose a surface

Quick start

Local Python tracking

pip install llmkit-sdk

from llmkit import tracked
from openai import OpenAI

costs = []
client = OpenAI(http_client=tracked(on_cost=costs.append))

client.chat.completions.create(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "Summarize this incident."}],
)

print(f"${sum(item.total_cost or 0 for item in costs):.6f}")

The transport reads provider usage metadata and estimates cost from the bundled pricing catalog. It does not send tracking data to LLMKit.

Zero-code CLI tracking

npx @f3d1/llmkit-cli -- python my_agent.py

Use -v for per-request output or --json for machine-readable results.

Gateway mode (existing key)

Gateway examples require an existing LLMKit API key. Account creation and key management are temporarily unavailable while the authenticated service is restored. If you do not already have a key, use one of the local tracking paths above.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.llmkit.sh/v1",
    api_key="llmk_your_key_here",
)

response = client.chat.completions.create(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "Draft a release note."}],
)

The budget path

The control path is built around three boundaries:

• Atomic admission: a Durable Object owns reservation state for each budget scope. Concurrent requests cannot spend the same remaining balance.
• Dispatch-aware idempotency: deterministic failures before dispatch release the key. After provider dispatch may have occurred, failures remain terminal to avoid duplicate spend.
• Bounded responses: non-streaming bodies and individual SSE frames have explicit byte limits. LLMKit cancels upstream reads when a limit is exceeded.

Request receipts bind the admission decision, provider attempt, settlement, and analytics handoff with stable identifiers. Database writes use an outbox, so an analytics outage does not silently erase budget evidence.

MCP server

{
  "mcpServers": {
    "llmkit": {
      "command": "npx",
      "args": ["-y", "@f3d1/llmkit-mcp-server"]
    }
  }
}

Five local tools inspect supported Claude Code sessions and Cline task data without an LLMKit key. Six gateway tools query spend, budgets, keys, sessions, and service health when LLMKIT_API_KEY contains an existing key. Together they expose 11 tools.

Pricing data

The pinned catalog is a bundled reference snapshot, not a live quote. One source file, packages/shared/pricing.json, records the snapshot date and generates the TypeScript, Python, and MCP tables. CI rejects drift between the source and generated files. The public site renders only populated provider tables and displays the source date.

The public comparison endpoint requires no account:

https://api.llmkit.sh/v1/pricing/compare?mode=text-token&models=anthropic%2Fclaude-sonnet-4-6%2Copenai%2Fgpt-4o&input=1000&output=1000&cacheRead=0&cacheWrite=0

The endpoint prices only the exact model keys supplied by the caller. It does not search for or recommend the cheapest model. Pricing is an estimate, not a provider invoice. Provider billing rules, model modality, and catalog freshness remain part of the error boundary.

Evidence and current boundary

See STAGING_PROOF.md for the isolated hosted proof contract. It deliberately refuses production targets and dirty worktrees.

Project policy and design

Development

git clone https://github.com/smigolsmigol/llmkit
cd llmkit
corepack [email protected] install --frozen-lockfile
corepack [email protected] build
corepack [email protected] quality:pr

Run the Worker locally with development-only bindings:

corepack [email protected] --filter @f3d1/llmkit-proxy dev

Generic deploy commands are intentionally omitted. Staging and production use separate guarded scripts with explicit target confirmation.

Security

Provider credentials are encrypted with AES-256-GCM using a random IV and owner/provider-bound additional authenticated data. LLMKit API keys are hashed before storage. CI includes secret scanning, static analysis, dependency review, CodeQL, and package provenance checks.

Read the security policy and architecture and the machine-readable Security Insights snapshot. Please report vulnerabilities through GitHub private vulnerability reporting or email [email protected].

License

MIT