Moli, Composio Instant, and Heal: Three Rooms to Put an Agent In
An agent demo dies at the same place every time, and it isn't the model. The agent has nowhere to stand.
0:00 / 6:52
Moli, Composio Instant, and Heal: Three Rooms to Put an Agent In
Three of last week's launches don't sell intelligence. They sell the room an agent works in: Moli is a browser it can drive, Composio Instant is a wallet it can spend from, and Heal is a running copy of your own product it can go and break 1. Each one rents the room and leaves you the fence around it.
The short version. Install Moli today if an agent has to read the web at volume and your pages aren't canvas or video — it is free, the install takes two seconds, and the only test that matters is whether your own pages survive at 82%. Take Composio Instant for a prototype that needs paid tools without ten signups, and write the spend cap yourself on day one. Book the Heal call for a release you are worried about, and don't plan a sprint around it until you have seen a price.
| Tool | Best first job | Setup friction | Time to first value | Small-team price | Integration pain | Practical pick |
|---|---|---|---|---|---|---|
| Moli by Lexmount | An agent that has to read, crawl or drive pages at volume | One-line shell installer, no account, no Chrome to download; the binary is 108 MB 2 | 2.2 seconds to install here; a plain page answered in 0.06 s at 50.8 MB, a JavaScript site in 0.44 s at 70.5 MB | Engine free under Apache 2.0 or MIT; the hosted service starts free with 10 browser hours a month, then $12/month for 100 hours 3 | 81.88% task success against Chrome's 99.85% on the vendor's own benchmark, and no canvas, WebGL or media playback 2 | A team paying for browser memory today, whose pages are ordinary DOM and forms |
| Composio Instant | A prototype that needs paid third-party tools without opening ten accounts | pip install composio took 5.7 seconds here and then stopped for an API key or a login; Instant is off by default and marked experimental 4 | A new wallet holds $2, so the first call is minutes away once you are in the dashboard 5 | Hobby is free (100,000 tool calls a month); Pro is $29/month with $29 of credit that expires monthly; provider usage at 0% markup 6 | No built-in dollar cap, and the prepaid balance is not a hard cap because balance checks are asynchronous 4 | A prototype or an internal tool where a human still watches the spend |
| Heal | Proving which bugs a release actually has | Connect GitHub; free trial with no card; local runs lean on Docker sandbox stacks 7 | A run builds a production-like sandbox first, so minutes to hours rather than seconds 8 | No price published; the trial needs no card and plans appear inside the app 9 | The benchmark is the vendor's own, two of its five codebases are theirs, and some findings are open product questions rather than bugs 8 | A team with a release coming and a repo it can point a robot at |
Moli: the browser, and the 82 percent
Moli launched in the week of October 19 and landed 13th on Product Hunt's weekly board 1. It comes from Lexmount, and it is a headless browser engine written from scratch in Rust rather than a wrapper around Chromium, published under Apache 2.0 or MIT at your option 2.
I ran their installer on this machine. It downloaded and installed in 2.2 seconds into
~/.local/bin, and left a 108 MB binary at version 1.1.19. Then I pointed it at pages: example.com came back as Markdown in 0.06 s with a peak of 50.8 MB of memory, and a JavaScript-heavy product site came back in 0.44 s at 70.5 MB. A screenshot needs --layout, which the error message tells you; with it, a 1920×1080 PNG took 0.33 s.Their own benchmark is the number that matters more than any of mine. On 1,308 comparable tasks, Moli passed 81.88% and Chrome passed 99.85%; median peak memory per task was 92 MiB against Chrome's 697, and median CPU 100 ms against 687 10. Read that as the vendor measuring itself: the run is theirs, and it was made on version 0.1.1, not the 1.1.19 I installed. The shape of the trade is still clear — roughly a seventh of the memory for roughly four-fifths of the tasks.
The limitations section is where the rest is written down. No pixel-for-pixel parity with Chrome, and no high-fidelity canvas, WebGL or media playback 2. If your agent has to watch a video, drive a 3D viewer or trust a screenshot of an animated chart, this is the wrong engine, whatever it saves you.
One test your agent will hit before any of that: I pointed Moli at a Product Hunt product page, and it came back with the security-verification interstitial in about a second — at 225.8 MB of memory, three times what a normal page cost. Light and fast is not the same as stealthy.
What it does give you is the integration surface, and it is genuinely boring in the way you want. It speaks the Chrome DevTools protocol: I started its server on port 9222 and asked it who it was, and it answered as Chrome/145.0.0.0 with a WebSocket debugger address. Existing Playwright and CDP code connects to it without a rewrite 11.
The hosted version is a separate decision from the engine. Lexmount's own plans start free with 10 browser hours and 1,000 fetches a month, then $12/month for 100 hours and 25 concurrent sessions, with CAPTCHA handling and persistent contexts held back for the paid tiers 3. If you self-host, the engine is free and the machine is yours.
Composio Instant: the wallet, and the missing cap
Composio Instant launched on Tuesday and took second place for the day 12. The pitch is one wallet instead of ten subscriptions: 1,500-plus tools, from search and scraping to transcription and image generation, called on Composio's provider accounts at 0% markup on top of what the provider charges 5.
I installed their Python SDK, version 0.30.0, in 5.7 seconds. The first session call stopped immediately:
ApiKeyNotProvidedError — either provide a key, export it, or run their login. That is the shape of the product rather than a bug in it: Composio is a hosted account, and the published Instant rate for a given tool only becomes visible once you are inside, because the pricing page loads those rates from the dashboard 6.The finding worth your time is in the documentation. Instant is off by default and marked experimental; you turn it on in project settings, and the docs tell you to contact support if the setting is missing 4.
There is also no dollar cap, and the docs are unusually direct about it. There is no built-in spend cap, the prepaid balance is not a hard cap because balance checks happen asynchronously, and
return_instant_charge reports costs after the call rather than authorising them. Their recommended way to bound it is four steps of application-side work — reserve an upper bound per call, reject what you cannot bound, bind the human approval to the exact tool and arguments, reconcile against the log afterwards 4.Prices are simple where they exist. Hobby is free with 100,000 tool calls and 50,000 trigger events a month; Pro is $29/month with $29 of usage credit that expires each billing month rather than rolling over, and higher rate limits 6. Instant provider charges come out of your balance separately, and their pricing page warns that rates may change 6.
What you buy is the removal of ten signing-up ceremonies from a prototype. What you keep is the budget, in code you write, with an approval step that can only live in your executor — the docs say plainly that handing the agent the MCP endpoint bypasses the gate and that agent instructions are not an enforcement boundary 4.
Heal: the running copy, and the price you can't see
Heal launched on Thursday the 22nd and came fourth for the day 13. The distinction it sells is narrow and real: a code reviewer reads the diff of a pull request, and Heal instead learns to run your application inside a sandbox that replicates production and goes looking for what breaks when a real person uses it 9.
The rule it holds itself to is the reason to care: what it cannot reproduce, it does not file. Every defect arrives with the screen it failed on, the expected and actual behaviour, and a failing test that turns green once the bug is fixed 9. It mocks or emulates payments, mail and third-party APIs, and it fires security scanners at the running sandbox, including ZAP, sqlmap and nuclei, keeping only what a test can reproduce 9.
Their benchmark, published this month, replays real pull-request history: five products, 236 pull requests, with CodeRabbit, Greptile and cubic installed on their default settings alongside Heal 8. Their finding is that 70% of the defects Heal filed were not found by the review agents, and that stacking reviewers adds little because they overlap heavily. On Formbricks, 56 pull requests produced 310 defects from Heal against 117 from all three reviewers combined.
The caveats are the vendor's own. Two of the five codebases are theirs, they did not assess whether the reviewers' flags were relevant while Heal's were reproducible by construction, and they say so in the limitations section 8. Their own summary also concedes that some behaviour issues are open product questions rather than bugs, which is the part a small team has to rule on itself.
Two practical walls. There is no published price:
heal.dev/pricing returns a not-found page, the site offers a free trial with no credit card, and plans are described inside the app once you connect a repository 9. And a local run needs somewhere to build: the changelog mentions sandbox Docker stacks and parallel lanes, so it wants Docker and a machine that can hold them 7. I could not run it here for exactly that reason, and that section of this episode is documentation rather than a test.What it returns is the part worth paying for: a bug you can hand to a coding agent as a task, with the test that proves it, instead of a comment asking you whether the line is reachable. Their own changelog shows the workflow closing — an MCP server so Claude can take a defect, fix it, and re-run the test 7.
The startup-job picks
For an agent that has to read the web at volume, install Moli today. It costs nothing, the install is two seconds, and the test is whether your own pages clear the 82% line — run your ten most important pages through it before you run anything else 2. If your pages are canvas, video or animation, stop now and pay for Chrome.
For a prototype that needs paid tools without ten signups, turn on Composio Instant and write the cap before the first call. Two dollars of credit and 0% markup make the trial honest; the unfenced balance against a rate that may change is what you have to engineer around 4.
For a release you are worried about, book the Heal demo and run it on one repository. Do not commit a sprint to it until you have seen a price and watched it build a sandbox once, because both of those are the parts you cannot read off the site 9.
The thread is the fence. Moli sells the speed and hands you the compatibility call. Composio sells the wallet and hands you the budget. Heal sells the proof and has not yet told you what it costs. Cheap rooms are the story of this year; the fence around them is still the part you build.
Two limits on what you have just read. This is a public-source buying screen rather than a logged-in trial: Moli was installed and run here, Composio's SDK was installed and stopped at its own login wall, and Heal's section rests on their pages and documentation because a local run needs Docker this machine doesn't have. And check the reversibility before anything sits in front of paying customers — especially the wallet, whose whole design assumes you keep the approval step.
References
- 1Product Hunt weekly leaderboard, week of October 19 2026
producthunt.com
- 2Moli repository
github.com
- 3Lexmount pricing
browser.lexmount.com
- 4Composio Instant Tools documentation
docs.composio.dev
- 5Composio Instant
composio.dev
- 6Composio pricing
composio.dev
- 7Heal changelog
heal.dev
- 8Heal benchmark, October 2026
heal.dev
- 9Heal
heal.dev
- 10Lexbench headless browser benchmark
github.com
- 11Lexmount cloud browser
browser.lexmount.com
- 12Product Hunt daily leaderboard, October 20 2026
producthunt.com
- 13Product Hunt daily leaderboard, October 22 2026
producthunt.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.