BYOAIK — Bring Your Own AI Key: a directory of AI tools that run on your own API key.

Agentic Ledger

Developer claimed — ownership proven 2026-08-10

✦ Most Innovative

Local-first flight recorder for AI agents: every call, tool use, and dollar captured by a transparent proxy on your own key.

Website Source code

Category:
Developer Infrastructure
Pricing:
Open Source
Open source:
Yes (MIT)
Self-hostable:
Yes
Local-first:
Yes
Platforms:
CLI, Self-hosted, Docker, Web
AI providers (bring your own key):
Anthropic, OpenAI, OpenRouter, Custom / OpenAI-compatible, LM Studio
API key storage:
Stored locally on device
Key risk level:
LOW
Trust score:
94/100

Agentic Ledger is an MIT-licensed, local-first observability proxy for AI agents. You point your agent's base URL at it (one env var: ANTHROPIC_BASE_URL or OPENAI_BASE_URL) and every LLM call is captured with full prompts, tool calls, and cache-aware cost accounting, with no code changes to the agent. It is explicitly bring-your-own-key by construction: the proxy passes your existing provider credentials straight through and never stores them, and keys it does need (for optional replay) can be supplied via file using the Docker-secrets _FILE pattern to keep them out of shell history. Team cards let one real provider key be shared as revocable per-team allowances with hard dollar budgets enforced in the request path. Data lives in a single SQLite file on your machine with no account, no telemetry, and no cloud component. Distributed via pip and a multi-arch Docker image, with a CLI and a local web dashboard; supports Anthropic and OpenAI wire formats, OpenAI-compatible gateways such as LiteLLM and OpenRouter, LM Studio for local replay, plus OTel ingest from Gemini CLI, Codex CLI, AutoGen and Pydantic AI.

Why we picked it: Most Innovative

The only tool here that makes an agent's spending legible and enforceable: a transparent proxy that captures every call and cost with no code changes, then puts hard dollar budgets in the request path where the agent itself cannot lift them.

  • Transparent proxy: one env var, zero code changes, works with any framework or provider
  • Cache-aware cost accounting, including the prompt-cache traffic most tools miss
  • Loop Lens for overnight agent runs: cost per iteration, stuck-loop detection, circuit breaker
  • Replay a whole run against a different model and get a match report card
  • Cost what-if: reprice a captured run on another model with no API calls or spend
  • Hard budgets enforced before the call reaches the provider, per session, agent, user, and day
  • Kill switch that survives proxy restarts and sits outside the agent's reach
  • Team cards: one real provider key, revocable per-team allowances with their own budgets
  • Flow and Trace views built from real parent links, plus OTel ingest
  • MCP server that returns compact summaries instead of dumping your context window

Disclosure: Agentic Ledger is built by the maintainer of BYOAIK. It is listed on the same published criteria as every other tool, and its trust score is unadjusted.

Why this trust score (94/100)

Trust measures how the tool treats your API key and how much of that has been verified. It contains no popularity signal.

  • Key Safety 23/25 — The key is held on your own device.
  • Request Routing 20/20 — Requests never leave your machine. The provider endpoint is configurable, so you can point it at the provider or your own gateway.
  • Transparency 20/20 — Source is public under MIT, so anyone can check how the key is handled. Key handling was located in the published source.
  • Privacy 14/15 — Local-first: it works without sending your data anywhere. Can be self-hosted, so the data path stays inside infrastructure you control.
  • Maintenance 10/10 — Actively developed — commits within the last three months.
  • Verification Confidence 7/10 — Key handling was found in the published source by BYOAIK's scanner.

What was checked

Verification tier: SOURCE_VERIFIED — derived from the evidence below, not set by hand.

  • [REPORTED · RESEARCH] This listing was compiled by an AI-assisted research pass over the tool's public website, README and documentation. No person independently confirmed it.
  • [STRONG · SOURCE_SCAN] The user supplies their own OpenAI key: the project references its API key variable. source
  • [STRONG · SOURCE_SCAN] Reads the API key from an environment variable at runtime. source
  • [STRONG · SOURCE_SCAN] Supports a local model backend, so it can run with no cloud provider key at all. source
  • [STRONG · SOURCE_SCAN] The provider endpoint is configurable, so requests can be pointed straight at the provider or at your own gateway. source
  • [CONFIRMED · SOURCE_SCAN] Ships a container definition, so it can be self-hosted on your own infrastructure. source
  • [CONFIRMED · SOURCE_SCAN] Most recent commit 2026-08-06 — about 0 month(s) ago. source
  • [REPORTED · DEVELOPER] The developer states the key is handled as LOCAL_DEVICE.
  • [REPORTED · DEVELOPER] The developer states requests are routed LOCAL_ONLY.

How your API key is handled

Your API key is supplied via an environment variable or local config file. Requests are sent directly to the AI provider. Because it can be self-hosted, your key never has to touch a third-party backend.

Setup

Install the tool, set your provider API key as an environment variable (or in its config file), and pick a model.