Public Beta Scope

Try the public beta

Open the public beta. No installation, signup or invitation is required. The browser compiles your editable AssemblyScript program, signs the request with its own temporary key, and sends it to a real remote Froglet node. The node runs Wasm under limits and returns signed evidence that the browser checks locally.

• Run the default input {"a":6,"b":7}: the program returns sum 13 and product 42.
• Change the code and input to run your own bounded computation.
• Query five invented terminology rows, including an ambiguous blood mapping.
• Try an infinite loop to inspect a signed execution-budget failure.
• Retry the saved job to recover its result; an accepted-job read does not create another deal.
• Download or copy the signed exchange into the standalone verifier.

The public provider runs immutable native release v0.4.7-beta.1 on Fly in Amsterdam, behind an allowlisted ingress and the website’s same-origin transport. On 6 October 2026, operator-run HTTPS checks exercised arithmetic, edited code, selected data, signed failure and recovery. A real browser also completed a fresh public request. These checks do not establish a fresh LLM session, independent human onboarding, sustained load capacity, scientific correctness or paid settlement.

Each Wasm job has at most 2 seconds, 8 MiB memory and 50 million fuel. There is one provider with finite shared allowances: 1,000 deals, 3,000 quotes and 2,000,000 milliseconds of reserved execution, plus persistent HTTP traffic ceilings. The page shows the observed remaining native allowance. Busy or exhausted capacity can refuse calls; this is a small free beta, not an uptime guarantee.

Use public examples only. Program bytes and input reach the provider and job records persist there. The browser saves the signed request in session storage for recovery, but does not save or export its private signing seed. Paid jobs, GPU work, arbitrary containers and hosting your own service are outside this public demo. For your own service, prepare a local setup.

Retired hosted trial

The older try.froglet.dev session gateway and its ai.froglet.dev origin were retired on 8 October 2026; session creation had returned a Cloudflare origin-DNS error since 24 September. Do not call their APIs or use the old prompt that starts with Read https://try.froglet.dev/llms.txt: they no longer produce Froglet evidence. The public beta above replaces that trial. The former contract remains in git history.

Only the first-party service is retired. Operators can still enable the node’s own hosted-trial routes for their deployments; see the FROGLET_HOSTED_TRIAL_* settings in the configuration reference.

Agent tasks

Agents choose a task from /agent-tasks.json. Without installing anything, an agent can run receipt-artifact-verify on signed records the user saved from the public beta, or marketplace-evidence to read the public marketplace snapshot. publish-service, consume-service and bounded-compute need a local node. The public beta is built for a browser and has no agent contract.

Moving to a local setup

An agent that can run tools should use this staged local flow:

1. Report the evidence it actually observed, if any: verified records the user pasted, marketplace reads, and mismatches with these docs.
2. Explain what that evidence does and does not show for your actual context. If it cannot see your repo, files, data, tools, or workflow notes, it should say that before making claims about your setup.
3. Ask for the local install profile before running commands:
   • agent: claude-code, codex, or manual for no-clone setup; openclaw only for source/plugin mode
   • footprint: auto (recommended native-first), native, docker, binary, or source
   • role: consumer, provider, or both
   • payment: none, lightning-lnd-rest, stripe-test, stripe-live, or x402
   • network: clearnet, tor, or dual
   • marketplace: none, https://marketplace.froglet.dev, or a custom URL
   • first use case: consume, publish, witness, hash-verify, notarize, paid compute, or prepare batch work later
4. Call plan_install with those choices. A complete call is non-mutating and returns one published immutable release tag, release-manifest/bootstrap hashes, persistent paths, process-manager impact, exact command preview, and install_approval_hash. Show the whole plan and wait for approval.
5. Only after approval, call get_install_guide with the exact returned release_tag and install_approval_hash plus the unchanged profile. The agent runs its verified-temp-file command through the host shell, not through Froglet runtime.
6. For a disposable-host proof from a trusted checkout, run bash scripts/fresh_host_quickstart_smoke.sh.
7. Verify provider health, runtime health, and local Froglet status.
8. Call plan_use_case before implementing the concrete workflow. For batch requests, start with existing async task status primitives and do not claim multi-item fan-out until Order 44 lands. GPU execution is currently refused for every runtime; an earlier Docker/T4 smoke covered a previous execution path and does not qualify GPU work now.

If your LLM cannot run tools

Some web chat interfaces cannot fetch URLs, make HTTP requests, run shell commands, or call MCP tools. In that case the honest answer is:

I cannot run Froglet from this chat interface. Open https://froglet.dev/services/ in a browser to try the public beta yourself, or use an HTTP-capable agent such as Codex or Claude Code for a local setup, and paste the outputs back here.

The fallback still has value: the model can explain what to run, what output to look for, and how to choose the local install profile, but it must not claim it observed live Froglet evidence.