I kept burning my weekly credits before the week was out. Every "token-saving" tool I tried claimed wins but none measured what I actually paid — real dollars, end to end, on real tasks. Token counts aren't a bill.

So I built LemonCrow: a runtime that lives inside your existing agent host, changes nothing about your workflow, and squeezes out the maximum saving it can. Every number below is an absolute-dollar measurement (BENCHMARKS.md) — not a token-count hand-wave.

Install from a checksummed GitHub release:

curl -fsSL https://github.com/lemoncrow-lab/lemoncrow/releases/latest/download/install.sh | bash

Then initialize it inside the project where you use your coding agent — no login, no network:

    cd your-project
    lc init  # Initializes your repo and index it.

ChatGPT chat usage doesn't burn API/agent-plan credits the way a coding agent's own usage-based billing does. So LemonCrow's tools (search, read, edit, bash) can be exposed through a tunnel and driven straight from a ChatGPT conversation instead — same grounded tool surface, no separate agent bill:

lc chatgpt serve

Prints a pairing code and, by default, an auto-launched cloudflared tunnel URL (installs cloudflared on first use if missing). In ChatGPT: Settings → Plugins → Browse Plugins → (next to search) + → Create, paste the printed MCP server URL, set Authentication to OAuth, and approve the browser prompt with the pairing code.

Full request/response traffic is logged locally per run (path printed at startup; credentials and tokens are redacted) so you can audit exactly what ChatGPT sent and got back.

Known ChatGPT-side quirk: Sometimes chatgpt looses the tool aceess on a new chat message conversation and without reattaching it can't access the tool. Workaround is branchoff the chat and then reattach the tool and continue with your message. Permissions: If it complains about permissions or asks to reconnect, check in the Setting -> Plugins, it has Allow All permission

⚠ The pairing code is a password — don't share the tunnel URL. This exposes shell-grade tool access (bash, edit) to this machine while the server runs. Stop it (Ctrl-C) when you're done.

LemonCrow records all sessions locally so you can inspect, audit, and debug exactly what your agent did.

lc session stats     # read-only report of wasted tool calls and round-trips
lc session replay    # replay a recorded session through the real LemonCrow tools

Both are local and read-only — no model re-run, nothing transmitted.

_{Replay recorded agent sessions locally with full tool visibility and resource usage breakdown.}

LemonCrow keeps your existing coding agent and changes the working set around it:

_{Your codebase's code universe — 28,462 symbols · 38,811 nodes · 23,894 calls. Live, local, on this repo.}

On Claude Code, lc init gives the agent five grounded tools and hides the equivalent built-ins — one way to do each job, not two. Other hosts use the strongest equivalent controls they expose.

What's unchanged: the host, the model, your workflow. Full internals: Architecture.

Packaged in integrations/agents/ — each a distinct capability grant (subagent name lemoncrow:<mode>):

Optional Packaged in integrations/skills/: /lemoncrow, /benchmark, /orchestrate, /swarm, /perf-review, /ux-review, /recall.

• It is not a hosted service. There is no cloud backend, dashboard account, or team collaboration server.
• It does not run your model for you — you bring and configure your own provider/API key (Anthropic, OpenAI, Ollama, …).
• It does not guarantee the benchmark deltas below on your repository; results vary by task, codebase, and model.
• Some integrations are early or in progress; behavior varies by host (e.g. session-close verification is enforced on Claude Code, advisory elsewhere).

• Runs locally. After install, all core functionality works offline and contacts no LemonCrow-controlled server.
• Telemetry is OFF by default and strictly opt-in. A global DO_NOT_TRACK=1 / LEMONCROW_TELEMETRY=off environment kill switch is honored.
• No identifier, repository path, source code, or symbol name leaves your machine. There is no crash reporting.
• The only network calls the runtime makes are ones you initiate (lc update, which checks GitHub Releases).
• Full detail: docs/privacy.md.

• Operating systems: Linux and macOS (primary); Windows is partially, never tested.
• Runtime: Python 3.12–3.13, managed with uv.
• Agent hosts: Claude Code, Codex and opencode today; Copilot, Cursor, Hermes, and Antigravity are in progress. Any MCP-compatible agent can connect to the same tools.
• Build requirements: uv, a C toolchain (only if you opt into the mypyc performance build; a pure-Python build works without it), and git.
• Known limitations: see What LemonCrow does not do and Troubleshooting.

These are fixed results from pinned benchmark runs — not a live counter. Every headline number links back to committed raw runs and methodology in BENCHMARKS.md. The model, tasks, containers, turn limits, and verification harness were held constant. Results are mixed by design and include a regression (SWE-bench Lite below).

_{Both arms 89 tasks x 5 reps = 445 trials on the same dataset (LemonCrow's Harbor run vs the Claude Code 2.1.205 leaderboard run), so correctness is directly comparable. LemonCrow also sends 91.8% fewer fresh input tokens (1.05M vs 12.87M); cost is not a matched comparison here — see .}

SWE-bench Verified detail (250 runs a side) — one-shot search collapses the grep-and-read loop, so turns, wall-clock, and tool calls drop together:

Indexing throughput and search quality hold up at repository sizes agents actually hit. A cold full rebuild of the Linux kernel core (1.24M symbols, 4.5M lines) and retrieval quality vs grep-class tools on ~7,200 query/answer pairs across 14 repos:

Ranked search is ~1.9x more accurate than ripgrep at a still-interactive p95; ripgrep wins raw latency but not what it finds. Per-repo indexing table and the full 13-tool retrieval comparison: BENCHMARKS.md.

Reproduce any of this from committed raw data: see BENCHMARKS.md and docs/benchmarks/results.md.

• Installation
• Troubleshooting
• Benchmarks · full results with methodology
• CLI reference
• Architecture
• Privacy & network behavior
• Maintenance-mode transition (audit & rationale)

Uninstall LemonCrow and its host integrations, preserving your data by default:

bash scripts/uninstall.sh

To also remove all LemonCrow-managed local state (databases, caches, logs, the local installation identifier, and configuration):

bash scripts/uninstall.sh --purge

The uninstaller stops background services, removes user-level systemd/launchd units, removes LemonCrow-owned host-integration entries (without touching unrelated agent-host configuration), reverts LemonCrow's PATH changes, and prints exactly what was removed and preserved. Preview with --dry-run.

If you want to build LemonCrow from source or run a local development setup, clone the repository and run the local installation script (see Installation for full details):

git clone https://github.com/lemoncrow-lab/lemoncrow
cd lemoncrow
bash scripts/local.sh

LemonCrow is free and open-source software under the Apache License, Version 2.0 lemoncrow.pro engine is planned to be released soon. See LICENSE and NOTICE.