Skip to content
An MCP server for Claude Code, Codex and Cursor. No API key.

One frontier brain.
An elastic workforce.

Keep your best model as the strategist. Leverage recruits local, free and connected models underneath it, runs independent work in parallel, replaces workers that fail and verifies every result.

Run a real mission nowSee it in action

No account, no key. The live page runs one bounded mission per visitor every ten minutes; every other run on this site is a recording, and each page says which it is.

MCP-native·Local-first·Zero-dollar mode·RocketRide execution
Mission LVR-f8f72d56
Finish the forge-app receipt splitting library so the whole exis
Verified
Paid spend
$0.00
Quality target
95
Workers
7
Workforce
  • Backend EngineerPool · best-freeREPLACED
  • Backend Engineerkodro-tutorREPLACED
  • Backend Engineerkodro-tutorREPLACED
  • Backend EngineerPool · best-freePASSED
Execution
  • 02:23.88Task completed: Implement the split calculation
  • 02:23.90Hired kodro-tutor as Integrator
  • 02:29.60"Implement the request handler" verified · quality 100
  • 02:29.60Task completed: Implement the request handler
Tasks verified
4/4
Proof checks
8/8
Cognitive handoffs
3
Actual paid inference
$0.00
The product

Plan. Hire. Verify.

Frontier intelligence earns its price on architecture, trade-offs and hard reasoning. It should not spend the same premium compute on repository search and mechanical edits. Your host states an outcome and a policy. Leverage does the rest in three stages, and each one below is shown with what the recorded run actually produced.

Worth the premium
  • System architecture
  • Difficult trade-offs
  • Security judgement
  • Final review
Not worth the premium
  • Searching every file for one symbol
  • Forty boilerplate test cases
  • Mechanical migrations and refactors
  • Retrying a formatter that failed
01 · Plan

A validated task graph.

A planner is untrusted. Its graph is a proposal until it survives validation: cycles, dangling edges, escaping paths and tasks with no check are rejected before anyone is hired.

  1. moneyPASSED
    Implement money helpers
    no dependencies, runs first · 3 attempts
  2. validatePASSED
    Implement receipt validation
    no dependencies, runs first · 2 attempts
  3. splitPASSED
    Implement the split calculation
    after money + validate · 1 attempt
  4. indexPASSED
    Implement the request handler
    after split · 1 attempt

4 tasks from mission LVR-f8f72d56. Tasks with no dependency run in parallel; nothing runs before every dependency has passed.

Open the plan
02 · Hire

A job market, not a router.

Every task becomes a job posting. Each reachable model is scored on task fit, measured reputation, context headroom, availability and latency, then the best eligible one is hired and the reasoning is shown.

Job · Implement money helpers
Candidates
22 scored
Max cost
max $0.00
Privacy
prefer-local
  • kodro-tutorHIRED
    local · 15 prior jobs
    0.66
  • qwen2.5-coder:3b
    local · 11 prior jobs
    0.65
  • kodro-fast
    local · 3 prior jobs
    0.64
  • Pool · best-freeINELIGIBLE
    Already attempted and failed on this task
    –

Policy runs before scoring. The struck row is a candidate barred after failing this task. A paid model under a $0 budget is removed the same way, never in the pool rather than out-ranked.

Read the auction
03 · Verify

Evidence, not confidence.

A task is complete when a compiler, a test runner or the filesystem says so. Model self-confidence is recorded separately and is the smallest term in the score.

  • src/money.js existsPASS
  • test/money.test.jsPASS
  • src/validate.js existsPASS
  • test/validate.test.jsPASS
  • src/split.js existsPASS
  • test/split.test.jsPASS
  • src/index.js existsPASS
  • test/index.test.jsPASS

ProofPack for mission LVR-f8f72d56: 8 of 8 checks passed, $0.00 paid, 149.6s elapsed.

Read the ProofPack
The demo · 2:22 · narrated

What it is. What it is for. How it works. Then watch it work.

What Leverage is, the use cases, the harness, a real mission running on this site, and the same product driven from a Claude chat through its connector: a mission on a real repository, three guards for the failures at the top of the OWASP list, planned by a model, verified by the repository's own tests, zero dollars paid.

Rendered with HyperFrames from motion/compositions/demo.html over real browser sessions · voiced with ElevenLabs from scripts/narrate-film.mjs · the 1:51 site walkthrough is at /motion/walkthrough.mp4 and the 69-second motion version at /motion/film.mp4
Cognitive handoff · nine seconds

A worker fails. The work does not.

A rate limit is an infrastructure failure, not a wrong answer. Leverage keeps what the worker understood, hires a different one, and carries on. That is why a 429 costs a worker rather than the project.

Rendered with HyperFrames from motion/index.html · 1920×1080 · every figure in it is from the recorded run
Play what it built

Not a screenshot. The actual build, running here.

The tests were written first. A mission hired the workers and every module was verified by running those tests. Quill's surface was given as a shell, the way a canvas is given to a game; its Node server is replaced here by a small browser shim so you can use it with your own model endpoint.

A Claude-style assistant: conversations, context window, rate limits, model adapter, settings, projects, attachments, streaming.
Connect any OpenAI-compatible endpoint
Your endpoint, key and conversations stay in this browser. There is no key of ours in the page: a public key would be an open relay.
Cognitive handoff

Replace the worker, not the project.

When a model hits a quota, times out or cannot solve the job, Leverage captures a compact checkpoint of what it understood: decisions, files touched, checks already passing, what is left. It hands that to a replacement. The work continues instead of restarting.

Worker released
auto/best-free
stopped with RATE_LIMIT
Checkpoint
196 tokens
from 453 of context
Context reduction
57%
counted from the text, 3.6 characters per token
Cognitive handoff · replay

Drive it yourself.

Not a recording. This replays the recorded mission’s own event log at its real elapsed timings. Scrub to any point and you are looking at what the scheduler was actually doing at that moment.

Stage00:01.88
Pool · best-free
checkpoint
replacement

Hired Pool · best-free as Backend Engineer

worker.hired
Event log · 1 of 10
  • 00:01.88worker.hiredHired Pool · best-free as Backend Engineer
gaps compressed for replay · timestamps are real
Live execution

Watch a worker fail and the work survive.

Every line to the right is a real event from mission LVR-f8f72d56, read from its append-only log. The rate limit is injected and labelled as such in the stream, so what you are watching is how the scheduler handled it.

Events recorded
129
Workers hired
7
Handoffs
3
Paid spend
$0.00
LVR-f8f72d56 · execution
  • 00:01.88worker.hiredHired Pool · best-free as Backend Engineer
  • 00:01.89worker.startedPool · best-free started work
  • 00:01.89provider.rate_limitPool · best-free: INJECTED 429 rate limit (deterministic fault injection, not a real provider error)
  • 00:01.89worker.failedPool · best-free failed: RATE_LIMIT · INJECTED 429 rate limit (deterministic fault injection, not a real provider error)
  • 00:01.89checkpoint.createdCheckpoint cp_e79faf1d95be: 196 tokens captured from 453 of context (57% smaller)
  • 00:01.89handoff.startedHanding task to a replacement worker with checkpoint cp_e79faf1d95be
  • 00:01.89auction.completedWinner: kodro-tutor · 100% task fit · 29% verified success over 15 prior jobs · local runtime, no spend
  • 00:01.89worker.hiredHired kodro-tutor as Backend Engineer
  • 00:01.89worker.startedkodro-tutor started work
  • 00:01.90auction.completedWinner: kodro-tutor · 100% task fit · 29% verified success over 15 prior jobs · local runtime, no spend · 1 candidate excluded by policy
  • 00:01.90worker.hiredHired kodro-tutor as Backend Engineer (resuming from cp_e79faf1d95be)
  • 00:01.90worker.startedkodro-tutor started work
  • 00:25.59proof.checksrc/validate.js exists: PASS · src/validate.js exists
  • 00:25.59proof.checktest/validate.test.js: FAIL · exit 1: 'test failed'
Built in the open

Every number on this page has a file.

No subscriber count, no revenue chart, no logo wall. What this repository has is a record: the missions it ran, the tests that hold its claims, a scale harness, a capability probe, and a RocketRide bill read from billing. Each figure below is read from its file when the page renders, and the file is named so you can open it.

5
Recorded missions
in this repository
19/19
Tasks verified
by a compiler or test runner
38/38
Proof checks passed
every one re-runnable
7
Workers replaced mid-task
without restarting the work
$0.00
Paid inference
across every run
Tests
97
cases in tests/, counted from the files · npm run test
tests/*.test.ts
Control plane at volume
100/100
tasks · 0 duplicate claims · 0 budget overshoots · 16 handoffs for 16 scripted outages · the paid candidate struck on all 116 auctions
demo/scale-run.json
Capability probe
8 of 22
models passed every probe · 1 partial · 13 failed · the failures are why handoffs exist
demo/proof/capability-probe.json
Hosted pool
13
model ids on the allowlist, each answered a real completion · 18 of 512 listed ids did
demo/evidence/pool-sweep.json
RocketRide pipeline
READY
one worker through staging.rocketride.ai · 19.4s · 2.3 credits
demo/evidence/rocketride-run.json

The bill, read from billing.

Cloud workers run as RocketRide pipelines, and pipelines consume credits. Each run reads billing.getCreditBalance before and after, so the cost is measured rather than estimated. When billing cannot answer, the UI shows unavailable instead of a guess.

One inference, one pipeline
2.3credits
4188.6 → 4186.3
openrouter/minimax/minimax-m2.7:free · 19.4s
demo/evidence/rocketride-run.json
Live run on the deployed site
12.0credits
4592.6 → 4580.6
LVR-783bade5 · 4/4 verified · 69s
demo/evidence/live-run-LVR-783bade5.sse
Live run, plan written by a model
15.8credits
4144.9 → 4129.1
LVR-2d4d56e0 · 3/3 verified · 121s · planned at 63s
demo/evidence/live-planned-run-LVR-2d4d56e0.sse
Recorded run, plan written by a model
24.6credits
4186.3 → 4161.7
LVR-31eacf88 · 3/3 verified · 135s
demo/planned-run.json
Execution fabric

RocketRide runs the work. Leverage decides what work should run.

Leverage never re-implements execution. It decides which model deserves a job, then hands that job to a RocketRide pipeline and refuses to accept the result until a compiler or a test suite agrees. Here is a recorded mission where that happened.

Workers via RocketRide
3 of 6
3 of 4 tasks finished this way
Their output
All verified
same checks as every other worker
Cognitive handoffs
2
context cut 39% and 45%
Actual paid inference
$0.00
4/4 tasks passed

A worker whose cost class is free executes as a RocketRide pipeline rather than a local call. In mission LVR-bda3ba68 that is how 3 of the 4 tasks were finished, and their output passed verification. Open it and read the event log yourself.

Run one yourself, nowOr open the recorded run
Your repository

Point it at a repository and a model writes the plan.

The benchmarks run committed plans so they measure the workforce, not the planner. This mission had none: a planner model read the repository and its tests, proposed the task graph, the compiler validated it, and every task was held to the test file that covers it, then the whole suite was run once more.

Planned by
Nemotron 3 Nano 30B via NVIDIA
free route · 23.9s to a validated graph
Tasks verified
3 of 3
3 test commands the plan named, all real runs
Whole suite
Green
npm test after the last task passed
Cost
$0.00
24.6 RocketRide credits, read from billing

Mission LVR-31eacf88 keeps the proposal exactly as the model wrote it, in the Plan panel, next to what each task was held to. The code the workers produced is committed under demo/output/greeter and reproduces the hash in every proof. On your machine it is one command: npm run mission -- --repo=/your/repo --goal="..."

Open the planned missionOr read how it was measured
The workforce

Names that did verified work here.

Not a logo wall. Every name below passed at least one verified task, or wrote the plan, in a recorded mission in this repository: 5 models across 24 hires, plus the model that wrote a task graph, with the record beside each. Cost class says how it was reached: local runtime, free route through the hosted pool, or the seat your subscription already pays for.

Pool · best-freefree8/9 verified
kodro-tutorlocal3/5 verified
OpenCode (your subscription)host1/3 verified
MiniMax M3 via OpenRouterfree4/4 verified
MiniMax M2.7 via OpenRouterfree3/3 verified
Nemotron 3 Nano 30B via NVIDIAplannerwrote the plan

Hosts it runs inside, through MCP: Claude Code · Codex · Cursor · Windsurf · Zed. Pause the row to read it.

One mission

One strategist. 7 workers. Nothing paid for.

This is the workforce from mission LVR-f8f72d56, the recorded run in this repository. Green passed verification, amber was replaced mid-task and handed its understanding to a successor.

Local runtime calls
3
Free cloud calls
3
Self-learning workforce

It learns which models are actually good.

Observed on the machine that ran the recorded missions: which models passed verification on real work, rather than a leaderboard someone else published. Rates are shrunk toward a neutral prior and carry their sample count, so nothing here can claim more than it has earned. The deployment’s own live roster is under Models in Mission Control.

Measured model performance for this installation
ModelSourceJobsVerifiedRateMedianConfidence
auto/best-freefree2721
76
26.3shigh
samantha-v0.1:latestlocal22
75
4.1slow
samantha-core:latestlocal22
75
4.1slow
auto/best-reasoningfree22
75
16.4slow
auto/best-fastfree22
75
16.1slow
gemma3:4blocal21
58
8.3slow
auto/best-codingfree42
56
62.5slow
kodro-tutor:latestlocal2210
48
27.5shigh
Bring your own intelligence

Connect what you already pay for.

Leverage discovers what it can reach and keeps it in sync. The best route is the one that needs no key at all: run Leverage inside the agent you already use, and it can hire that agent’s own model through MCP sampling.

No API key. Your seat.
Claude Code

Runs Leverage as an MCP server and lends its own model back as a worker.

No API key. Your seat.
Codex

Same MCP sampling channel. Your seat, your rate limit, no key handed over.

No API key. Your seat.
Cursor · Windsurf · Zed

Any MCP host that offers the sampling capability can be hired from.

No API key. Your machine.
Ollama

Every model you have pulled, discovered automatically. Nothing leaves the machine.

Your key, encrypted
OpenAI-compatible

LM Studio, vLLM, llama.cpp, OpenRouter, a gateway of your own. One base URL.

Your key, encrypted
RocketRide

The execution fabric. Worker pipelines, traces and token accounting.

What we will not do

We will not drive a logged-in browser session to borrow a consumer subscription. ChatGPT Plus and Claude Pro have no API, and anything that claims to “connect” one is automating a web UI against its terms with your credentials. MCP sampling gets you the same model, legitimately, and Leverage never sees a password or a token.

See it build something you can play.

A gravity-arena prototype whose entire logic was written by free and subscription-backed models under a hard $0 budget, shipped beside the ProofPack that produced it.

Open the demo
Zero-dollar mode

When the budget says zero, zero means zero.

$0.00
Actual paid inference, recorded run
  • MIT licence. No account, no API key, nothing to download.
  • Five MCP tools. Your host stays the strategist.
  • Paid providers struck before scoring, asserted in the tests.
  • RocketRide credits read from billing, per run, never estimated.
Paid providersBLOCKED
Your host seatREADY
Local runtimeREADY
Free cloud routesREADY
Blocked paid attempts0
Questions

Answered from the files, not the brief.

Each answer points at the test, the run or the document that backs it. If one stops being true, the copy is what is wrong.

Do I need an API key or an account?

No. Leverage hires from the seat your MCP host already has, through sampling; from the agent CLIs already signed in on your PATH; from Ollama; and from free routes. The live page on this site needs nothing at all. A RocketRide key is optional on your own machine: with it, cloud workers run as pipelines; without it they are called directly, and the mission log says which happened.

Is $0.00 real, or a rounded estimate?

It is the settled paid spend from the run’s own ledger. A paid model under a $0 budget is removed from the auction before scoring, and the budget governor reserves headroom atomically before any paid call, so four concurrent workers cannot each see the last five cents and all proceed. Both are asserted in tests/invariants.test.ts. The one estimate on the site, frontier-equivalent cost, is labelled estimated everywhere it appears.

Does RocketRide actually do the work?

Every worker whose cost class is free runs as a RocketRide pipeline: webhook → llm_openai_api → response, with the auction’s chosen model patched into the LLM node at deploy time. That is how model selection becomes model execution. In mission LVR-bda3ba68, 3 of 6 workers ran that way and finished 3 of the 4 tasks. In LVR-719a8c22 every worker did, through the hosted pool. Credits are read from billing before and after each run.

The rate limit in the recorded run, is it real?

Injected, and labelled INJECTED in the event stream, in Mission Control and in the docs, because a demonstration of recovery should not depend on luck. A genuine 429 from pool:auto/coding:free appears independently in the capability probe, and two genuine test failures in the same run were handed off the same way.

Will it use my ChatGPT Plus or Claude Pro subscription?

Not by driving a logged-in browser session. Those plans have no API, and automating their web UI is against their terms and puts your credentials in the loop. Inside Claude Code, Codex or Cursor, Leverage borrows the host’s own model through MCP sampling: the same model, legitimately, with no password or token ever handled.

Can I point it at my own repository?

Yes: npm run mission -- --repo=/your/repo --goal="...". A planner model reads the repository and its tests and proposes the task graph; the compiler rejects cycles, dangling edges, escaping paths and tasks with no check, and a rejected plan fails the mission with the reason rather than falling back silently. Mission LVR-31eacf88 on this site was planned that way.

What has it been proven on?

Every number on this page comes from a mission that ran inside this deployment: JavaScript and TypeScript repositories with their own test suites, a mission RocketRide executed, a mission a hosted pool served, and a mission whose plan a model wrote. Small local models fail often, and the handoff machinery exists because they do: every replacement recorded here happened during a real run, with the checkpoint and the resumed task in the log.

Install

Give your host a workforce. Pick your host.

Five tools, not forty. A strategist needs to state an outcome, watch it, stop it, and inspect the evidence; everything else is Leverage’s job. The server talks to a running Leverage instance, so start one with npm run dev first.

Claude Code

One command. If the host offers sampling, its own model is hired back as a worker.

claude mcp add leverage -- node /abs/path/to/leverage/mcp/server.ts

Codex

Add the server to the host config. Same tools, same sampling channel.

{
  "mcpServers": {
    "leverage": {
      "command": "node",
      "args": ["/abs/path/to/leverage/mcp/server.ts"],
      "env": { "LEVERAGE_API_URL": "http://localhost:3000" }
    }
  }
}

Cursor · Windsurf · Zed

Any host that speaks MCP. Any host that also offers sampling can lend its seat.

{
  "mcpServers": {
    "leverage": {
      "command": "node",
      "args": ["/abs/path/to/leverage/mcp/server.ts"],
      "env": { "LEVERAGE_API_URL": "http://localhost:3000" }
    }
  }
}
Then, inside the host
Use Leverage. Finish this application. Budget $0. Quality production.
  • leverage_run
  • leverage_status
  • leverage_cancel
  • leverage_proof
  • leverage_models

Nothing to download here, and nothing to sign up for. If you would rather press a button than install, the live page runs a bounded mission inside this deployment.

Run a real mission nowMCP docs

Give your best model a workforce.

Run a real mission nowInstall the MCP server