rpcprobe

rpcprobe

which public ethereum rpcs answer today, and what they let you do. every endpoint that is listed somewhere as "no key needed" gets the same dozen small json-rpc requests: the chain id, five pings, log queries of growing width, one archive read, a batch, 1,024 blocks of fee history, the blob fee, eth_config, and whether a web page may read the answer. a table, json, and a dataset that fills itself every night.

$ rpcprobe
public ethereum rpcs, checked 2026-09-10 14:21 utc from this machine: 14 of 21 answer, head 25,947,555

endpoint     p50 ms  lag    logs archive batch receipts fee hist blob fee eth_config browser  client
tenderly-gw     238    1  10,000     yes   yes      yes     1024      yes         no     yes  tenderly 1.0
nownodes        246    0      10      no   yes      yes     1024      yes        yes     yes  geth 1.17.5
publicnode      246    0      50      no   yes      yes     1024      yes        yes     yes  geth 1.17.1
tenderly        250    1  10,000     yes   yes      yes     1024      yes         no     yes  tenderly 1.0
drpc            253    0  10,000     yes   yes      yes     1024       no         no     yes  geth 10.0.0
nodies          273    1      50     yes   yes      yes     1024      yes        yes     yes  reth 2.4.1
blockrazor      311    0      10      no   yes      yes     1024      yes        yes     yes  geth 1.17.5
blastapi        343    0      10     yes   yes       no     1024       no         no     yes  reth 2.4.0
mevblocker      348    0  10,000     yes   yes      yes     1024      yes         no     yes  mevblocker
merkle          358    ?       -      no    no       no        -       no         no     yes
zan             361    0       -      no   yes       no     1024      yes         no     yes  geth 1.17.5
1rpc            370    1       -      no   yes       no        1      yes        yes     yes  geth 1.16.7
meowrpc         401    0       -      no    no      yes     1024       no         no     yes
flashbots       550    0  10,000      no   yes      yes     1024      yes         no     yes  reth 1.11.2

no answer, or not without an account:
  onfinality   http 429
  blockpi      http 521
  cloudflare   Cannot fulfill request
  llamarpc     http 525
  ankr         Unauthorized: You must authenticate your request with an API key.
  omniatech    http 521
  gatewayfm    http 503

p50: median of five eth_blockNumber round trips from here. lag: blocks behind the highest head, read at one moment.
logs: the widest range of blocks answered for one address and one event (tried 10, 50, 100, 1,000, 10,000).
archive: an address's balance at block 1,000,000. receipts: eth_getBlockReceipts for one block.
browser: the answer carries access-control-allow-origin.

fourteen of twenty-one at 14:21 utc on 10 september, twelve half an hour earlier: nownodes and merkle come and go. every script i publish carries a fallback list of endpoints, and this is how that list gets made: tenderly, drpc and mevblocker answer a 10,000 block log query and know old state; publicnode is among the fastest but stops at 50 blocks of logs and keeps no archive; ankr wants an account now; cloudflare's and llamarpc's endpoints, still in a lot of tutorials, did not answer at all.

install

pipx install git+https://github.com/alinaschanz/rpcprobe

or clone it and run python -m rpcprobe from the folder. python 3.10 or newer, no dependencies.

use

rpcprobe                                   # every built-in endpoint, three at a time, about a minute
rpcprobe --only publicnode,drpc,tenderly   # just these (--list shows the names)
rpcprobe --rpc http://localhost:8545       # your own node next to them
rpcprobe --json                            # everything, including why a check said no
rpcprobe --csv today.csv
rpcprobe --summary-append data/daily.csv --quiet   # one row per endpoint per day, the dataset

what is checked

each endpoint gets about fifteen small requests, three endpoints at a time. when a check says no, the json says why, in the endpoint's own words (Block range too large: maximum allowed is 50 blocks, historical state is not available).

reading the table

• latency is from wherever this runs. from a laptop in berlin it is not what a server in virginia sees; compare endpoints within one run, and days within the dataset.
• logs 50 is not "broken": most scripts ask for 25 to 50 blocks at a time on a public node anyway. - means even 10 blocks were refused, usually a rate limit on that method.
• a rate limit (429) today can be fine tomorrow. the dataset is there so that one bad hour does not decide anything.

the dataset

data/daily.csv gets one row per endpoint every night at 00:53 utc, from a github runner, so the latency column compares from day to day. columns are in data/README.md; a rerun on the same day replaces the day.

exit codes and scripting

0 when at least one endpoint answered, 2 when none did or a flag was wrong. rpcprobe --json | jq '.endpoints[] | select(.ok and .archive and .logs_range >= 1000) | .url' is a fallback list for a script that reads history.

see also

• blobwatch, gasweek, bigmoves: the scripts that live on these endpoints
• the notes: alinaschanz.life, the short version on x

verify a release

every release carries the sdist and the wheel, a SHA256SUMS file, an opentimestamps proof of that file, and a build provenance attestation made in github's own signing flow. with the files downloaded into one folder:

sha256sum -c SHA256SUMS
gh attestation verify ./*.whl --owner alinaschanz
ots verify SHA256SUMS.ots

the commit itself is signed.

license

mit. the endpoints belong to the people who run them; this only asks them simple questions.