LemonCrow Runtime

Keep your coding agent sharp on real codebases

LemonCrow runs underneath Claude Code, Codex, and other supported hosts with a local code graph, exact-range reads, bounded output, durable memory, and verified runtime controls. On matched SWE-bench Verified runs: +12.0pp resolved, 37.7% fewer turns, and 23.7% faster — same model.

Matched proof, same model: 92.8% vs 80.8% resolved · 37.7% fewer turns · 23.7% faster · raw runs published · see live savings

Install · Working set · Results · Code-search benchmark · Pricing

Click for the full walkthrough on lemoncrow.com.

Install in 30 seconds

Run this once:

curl -fsSL https://install.lemoncrow.com | bash

Then create or sign in to a free LemonCrow account and activate it inside the project where you use your coding agent:

cd your-project
lc account login
lc init

LemonCrow wires in better tools behind the scenes and starts tracking local savings as sessions finish. Remote telemetry is on by default; turn it off anytime with lc telemetry remote off.

lc update

Check that everything is connected:

lc doctor

Account boundary: anonymous evaluation includes up to $50 in measured
savings over a rolling 30-day window. Creating or signing into a Free
account is required for uncapped core usage. Recall, handovers, verification,
and local multi-worktree swarm are included in Free.

The clean working set

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—not the raw transcript forever.
Verify	Notice code changes without tests or checks, then nudge the agent before it declares completion.

Open the existing dashboard with lc dashboard open, then choose Map. The
browser opens the full tracked source map: files, indexed symbols, and only
uniquely resolved call edges. Groups stay spatially grouped; filters split
source, tests, Markdown/docs, config, data, assets, and languages. Search any
symbol, click it for its exact source range, or double-click to focus callers
and callees. Live mode adds glow, edge flow, and camera-follow from the local
run ledger; switching it off leaves the same zoomable, clickable graph in
place. Raw source, diffs, stdout, and stderr never enter the map feed.

LemonCrow source map: a full repository call graph with 28,462 indexed symbols, 10,349 tracked files, 38,811 map nodes, and 23,894 resolved calls, one function focused to show its callers and callees.

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.

Find things in one shot -- no wandering the codebase call after call.

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: Claude Code itself, the model, your workflow. Full internals: Architecture.

Why a runtime, not just tools

A bare MCP server is a library the model can call if it remembers to. A runtime decides what's callable at all — four jobs, four layers:

Layer	Without it	With it
Agents — process isolation	a "read-only" agent can still edit	`explore`/`plan`/`research`/`review` hard-deny `edit`/`write` at the host-config level
Skills — standard library	multi-step procedures re-improvised every session	encoded once, invoked the same way every time
Hooks — interrupts	wasteful re-reads; "done" without a check	bad calls denied before they run; on Claude Code, session close blocked until verification ran (advisory on other hosts today)
MCP tools — syscall surface	agents fall back to grep-and-read under pressure	natives hidden — LemonCrow tools are the only surface for those jobs

The engine underneath — not just for code

Under the coding runtime is a domain-neutral context engine with one contract: one-shot retrieval — a single call returns what the model needs, packed to a token budget; follow-up queries are allowed, retry loops are never required. Retrieval, compression, memory, and replay do not assume code — code is the first conforming retriever (Retriever protocol, lemoncrow.core.capabilities.retrieval), chosen because the pain is mature and the results are measurable. The same engine is being opened to teams building any LLM agent: lemoncrow.com/engine.

Agents

Packaged in integrations/agents/ — each a distinct capability grant (subagent name lemoncrow:<mode>), not a persona:

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

Packaged in integrations/skills/:

Skill	What it does
`lc`	manage LemonCrow itself via the CLI
`benchmark`	measure savings on your repo — offline scan or live A/B
`orchestrate`	one structured multi-step task, routed to the right surface
`swarm`	N parallel attempts in isolated worktrees — best result wins
`perf-review`	gate a change on measured performance, not read code
`ux-review`	gate shipped UI on objective checks in a real browser
`recall`	retrieve what past sessions learned

Check your own savings

After LemonCrow is installed, inspect all savings and missed opportunities in
your local session history:

lemoncrow session stats

The report is read-only. It counts wasted tool calls and round-trips in your
Claude Code, Codex, and opencode sessions, then shows the associated token,
cost, and time opportunity.

Replay a past session

lemoncrow session replay plays back a recorded session and, for each native call, runs the real LemonCrow tool that would have replaced it — grep/read loops collapse into one code_search, whole-file reads become budgeted outlines, bash logs get capped. It then estimates — from that session alone — the cost, savings opportunity, and time LemonCrow would have saved. No model re-run, nothing written; opens a shareable HTML page.

lemoncrow session replay

Works on Claude Code, Codex, and opencode sessions. The saving is an estimate; the live re-measured A/B is lc benchmark local.

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. Community telemetry remains separate:
see live savings.

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	-	-	-	$8.68	$6.18	28.8% cheaper‡
Terminal-Bench 2.1, 89 tasks vs public leaderboard*	78.9% expected	78.7%	-0.2 pp	$96.76	$69.52†	28.1% cheaper†

_{* LemonCrow 1 rep/task vs public leaderboard 5-rep average. † 5 timed-out tasks excluded from cost. ‡ Caveman (free terse-persona system prompt, no install) also ran in the same invocation: $8.57, 1.2% cheaper than baseline -- LemonCrow's edge is durable cost, not just shorter replies; see /vs/caveman.}

LemonCrow vs baseline: dollars saved per run against baseline task cost, across SWE-bench Verified/Lite/Pro, exploration, Telegraphic Q&A, and Terminal-Bench

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
Bash	3,327	1,785	-46.3%
Read	1,733	1,050	-39.4%
Edit + Write	1,628	759	-53.4%
Search (code_search)	-	568	lemoncrow-only

Exploration detail (7 large repos × 5 reps, read-only Q&A, no edits):

Tool	Baseline calls	LemonCrow calls	Delta
Read	672	23	-96.6%
Bash	508	71	-86.0%
Search (code_search)	-	23	lemoncrow-only
Agent + orchestration calls*	79	1	-98.7%
Total tool calls	1,259	118	-90.6%
input	286,191	205,967	-28.0%
cache read	35,862,919	2,753,393	-92.3%
cache write	2,811,356	233,381	-91.7%
output	426,367	68,893	-83.8%
input + cache write	3,097,547	439,348	-85.8%

Source: exploration_2026_06_29 · telegraphic_2026_07_16.

One-shot code search vs 10 named tools

The search is the engine: right code in front of the agent on the first try. MRR and first-hit rate (rec@1) across ~7,200 query/gold pairs on 14 repos — 10 tools scored on the identical corpus. LemonCrow semantic search returns the right code first on 65.0% of queries at 390ms p95; the lexical path returns it first on 58.2% at 134ms p95:

Provider	MRR	rec@1	p95
LemonCrow +semantic (BGE)	0.727	0.650	390ms
LemonCrow lexical (default)	0.676	0.582	134ms
cocoindex-code (best rival)	0.557	0.457	595ms
serena	0.401	0.359	3834ms
ripgrep	0.376	0.320	66ms
universal-ctags (worst rival)	0.237	0.226	1ms

No one had scored these 10 tools against each other on a shared query set before -- each publishes its own number, on its own terms, against its own baseline.

What you get

Works with Claude Code, Codex, Copilot, Copilot CLI, and opencode today; Cursor, Hermes Agent, and Antigravity integrations are in progress. Any MCP-compatible agent (LangChain, the OpenAI SDK, Gemini ADK, ...) can connect to the same tools.
Runs locally by default.
Open-source runtime (Apache-2.0); the compiled engine (lemoncrow.pro) is proprietary and required at runtime.
Anonymous evaluation is capped at $50 in measured savings; a free account unlocks uncapped core usage.
Live local stats for cost, tokens, and savings; anonymous remote telemetry is on by default (opt out anytime).
Free: grounded local tools, recall, handovers, verification, and swarm.
Pro: larger-repo indexing, cross-vendor memory, compression, optimization, and reusable knowledge for $20/month or $200/year.

Learn more

Star History

License

Open-core. All source published in this repository is Apache-2.0 — free to read, use, modify, and redistribute.

The engine — the lemoncrow.pro package (the retrieval-quality and token-savings core) — is proprietary: it ships only as compiled binaries, its source is not published, and it is required at runtime (without it, LemonCrow falls back to a degraded built-in-only mode). See LICENSE for the exact terms.