Why I built this

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.

Quick start

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.

Or code from ChatGPT instead — free

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.

Flag Effect
--no-tunnel Bring your own tunnel (named cloudflared tunnel, ngrok).
--persistent --hostname X Stable URL via a Cloudflare named tunnel (needs a domain in your Cloudflare DNS); survives restarts instead of rotating.
--no-auth Serve/mcp with no authentication — the tunnel URL alone grants access. Prefer OAuth (default).

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.

Or Inspect your past sessions (Offline Replay, dry mode) before trying

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.

What LemonCrow does

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

Stage Runtime behavior
Find Rank symbols, definitions, callers, callees, usages, and exact source ranges before broad file exploration.
Read Return an outline or only the requested lines; cap noisy command and web output with recoverable spill files.
Carry Preserve useful task state through memory, deduplication, compaction manifests, and handover packets.
Verify Notice code changes without tests or checks, then nudge the agent before it declares completion.

What actually gets replaced

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.

LemonCrow tool Replaces (hidden from the model) Why
code_search Grep, Glob One call returns the symbol, its callers/callees, and ranked source — no grep-loop-then-read-whole-file. Ranked by call-graph centrality over a tree-sitter symbol table
read Read Returns an outline or the exact:L10-L40 range, budgeted, instead of the full file
edit Edit, Write Verified, cross-file edits in one call instead of per-file patch-or-create guessing
bash Bash Output is capped and structured so a noisy build log can't blow the context window
web_fetch WebFetch Strips a page to clean Markdown instead of a raw HTML dump

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

Agents

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

Agent Writes? Use
code Yes default interactive — edits, refactors, features
auto Yes fully autonomous — CI/headless runs
solve Yes end-to-end solving of a well-defined task
execute Yes one verified pass of an accepted plan
general Yes catch-all for mixed work
bare Yes minimal toolset, same discipline
explore No read-only exploration — locate and cite
plan No read-only planning, stops for human checkpoint
review No adversarial read-only review
research No external web research — cited memo

Skills

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

Code hygiene

The best code is the code you never wrote. Every LemonCrow persona climbs a fixed ladder before writing anything, stopping at the first rung that holds:

# Rung Do
1 Need it at all? Skip what the task doesn't require (YAGNI).
2 Already here? Reuse the helper, util, or pattern already in the repo.
3 Stdlib? Use the standard library before rolling your own.
4 Native feature? Reach for the platform capability that already exists.
5 Installed dep? Solve it with a dependency already in the tree.
6 One line? If it collapses to one line, make it one line.
7 Otherwise Write the minimum new code that works.

The ladder runs after the agent understands the problem, not instead of it, and it never trades away validation, error handling, security, or accessibility. Lazy about the solution, never about reading the code first.

When a change deliberately cuts a corner with a known ceiling (a global lock, an O(n²) scan, a naive heuristic), the agent leaves an lc-debt: <ceiling>; <upgrade path> marker. Harvest them any time into a ledger — lc debt flags any marker that names no upgrade path (no-trigger), the ones that silently rot:

lc debt          # ceiling + upgrade per deferred simplification, no-trigger flagged
lc debt --json   # same, machine-readable

The packaged code-audit workflow (Claude Code) adds an over-engineering lens that returns a concrete delete-list — code to remove, not rewrite.

What LemonCrow does not do

  • 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).

Privacy and network behavior

  • 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.

Supported environments

  • 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.

Results

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).

Benchmark Baseline correct LemonCrow correct Correct delta Baseline cost LemonCrow cost Cost delta
SWE-bench Verified, 50 tasks x 5 reps 80.8% 92.8% +12.0 pp $234.84 $165.45 29.5% cheaper
SWE-bench Lite, 10 tasks x 5 reps 98.0% 96.0% -2.0 pp $19.83 $17.51 11.7% cheaper
SWE-bench Pro, 10 tasks x 5 reps 88.0% 90.0% +2.0 pp $39.01 $30.61 21.5% cheaper
Exploration tasks across 7 large repos x 5 reps - - - $19.11 $6.29 67% cheaper
Telegraphic Q&A, 20 prompts x 5 reps - - - $8.40 $4.48 46.7% cheaper
Terminal-Bench 2.1, 89 tasks x 5 reps (matched)* 78.9% 80.0% +1.1 pp — — —

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:

Metric Baseline LemonCrow Delta
Turns 6,962 4,336 37.7% fewer
Wall-clock 14.3h 10.9h 23.7% faster
Total tool calls 6,700 4,167 -37.8%
Output tokens 3.04M 2.19M 27.9% fewer

Scale

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:

Metric LemonCrow Grep-class baseline
Linux cold index, lexical (1.24M symbols) 179s (~3 min) —
Linux cold index, zoekt trigram 13.7s —
Retrieval MRR (higher = better) 0.727 semantic / 0.676 lexical 0.376 (ripgrep)
Query latency, p95 134ms lexical / 390ms semantic 66ms (ripgrep)

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.

Learn more

Removal

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.

Development & Building from Source

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

License

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.