Triaging a JSON-lines app log on your own computer: lnav, Kelora, DuckDB and jq

lnav is our pick for checking the same application’s logs again and again, because it keeps broken lines in view and countable. Kelora is our Rising Star if you want to count and filter without a format file; for a one-off file, DuckDB or jq needs no setup. The lead is narrow.

PRACTICAL PICKSPreferredPickOur Preferred PickRepeated triage of one application’s logs$0

lnav

A free terminal log viewer that counts errors by endpoint in SQL, steps through one request’s messages and flags a broken JSON line as invalid instead of dropping it.

Narrow lead: 3.80 against 3.68 for DuckDB. Your log needs a format file first, and lines that do not start with { are not flagged.

Why lnav is our Pick
PRACTICAL PICKSRisingStarOur Rising StarCounting and filtering without a format file$0

Kelora

A free command-line tool, first released in May 2025, that reads JSON lines with no format file, counts the values of any field and filters on a request ID.

It skips a broken line with a message and still exits with success, and one developer writes it with AI agents and does not review the code.

Why Kelora is our Rising Star

Who is this article for? Developers with a JSON-lines log from their own application on their computer, who need to find the failing endpoint, count its errors and read one request’s records without uploading the log.

About 31 min read

What Preferred Pick, Rising Star and Ready to Rock mean

The quick list

All four are free. Compare what each needs before the first query, what it does with a broken line and how you count errors by endpoint.

All four are free, open-source tools with no account or subscription, checked October 20 and November 11, 2026. Your costs are setup, learning the query language and checking the results. Licenses: BSD-2-Clause for lnav; MIT for Kelora, DuckDB and jq.

The comparison at a glance

Compare what each tool does with a line that is not valid JSON, what it needs before the first query and how you count errors by endpoint.

Order follows our verdict for triaging a JSON-lines application log on your own computer: the Pick, the Rising Star, then the alternatives, not speed or popularity. All tools remain selectable.

Local tools for triaging a JSON-lines application log
FeaturelnavKeloraDuckDB CLIjq
Best fitPreferred Pick: repeated triage of the same application’s JSON-lines logsRising Star: counting and filtering JSON lines without a format fileAlso consider: a one-off file, if you write SQLAlso consider: a one-off file, if jq is already in your toolkit
Documented default on a line that is not valid JSONGiven an “invalid” level and highlighted; a line that does not start with { is shown as text, without that flagSkipped with a message on standard error, and the command exits with success; --strict stops at the first bad line; -l can leave a bad line uncountedThe query fails (ignore_errors is off); turning ignore_errors on skips such lines without listing themParse error giving the line and column, and the command fails; -R with fromjson? drops such lines silently
Before the first query on your own logWrite a JSON format file once, unless a built-in format (Bunyan, Pino, journald and others) matchesNone: JSON is recognized without a format file; a new option syntax and a small scripting language to learnNone; columns inferred from the first 20,480 recordsNone
Counting errors by endpointSQL GROUP BY on the format’s table, inside the viewer or with -n -c--freq path counts the values of a field; filter on the level firstSQL: SELECT … GROUP BY on read_ndjsongroup_by over the whole input, or paths piped to sort | uniq -c
InstallStatic downloads for Linux, macOS and Windows; Homebrew, Snap, FreeBSD; distribution packages can be much olderDownloads for Windows, macOS (Intel and Apple silicon) and Linux; Homebrew tap, cargo; .deb and .rpmOne program for Linux, macOS and Windows; Homebrew or winget; Windows needs the Visual C++ RedistributableDownloads for Linux, macOS and Windows (Windows on ARM added in 1.8.2), Docker image, most package managers
Maintenance and ownershipActive; mainly its creator, Tim Stack; crash past about 134 million lines fixed only in the 0.15 betaOne spare-time developer, best-effort support; AI agents write all of its code and its author does not review it; breaking changes in 2.1 and 2.2DuckDB Foundation holds the project; Amazon agreed in August 2026 to acquire DuckLabs, its developerjqlang organization; regular commits through October 2026 (checked October 20)
Reader reviewsNo reviews yet. Write a reviewNo reviews yet. Write a reviewNo reviews yet. Write a reviewNo reviews yet. Write a review

For JSON-lines logs from your own application, triaged on your own computer. Count broken lines before you trust a total: none of these tools is an unattended validator. All four are free. Public-source research checked October 20, 2026 and, for releases, licenses and security pages, November 11, 2026; no tool was run for this review, and the October 3 check is a dated historical record. Next substantive review: December 19, 2026, or sooner after lnav 0.15.0 or Kelora’s next release. Earlier check record.

All features and sources 13 criteria

Open the groups that matter to you. They compare the same selected tools. “Planned” means future work, not an available feature. Sources include the published conditions.

Triage and broken lines

Show 7 criteria
Triage and broken lines
FeaturelnavKeloraDuckDB CLIjq
Best fitPreferred Pick: repeated triage of the same application’s JSON-lines logsRising Star: counting and filtering JSON lines without a format fileAlso consider: a one-off file, if you write SQLAlso consider: a one-off file, if jq is already in your toolkit
Documented default on a line that is not valid JSONGiven an “invalid” level and highlighted; a line that does not start with { is shown as text, without that flagSkipped with a message on standard error, and the command exits with success; --strict stops at the first bad line; -l can leave a bad line uncountedThe query fails (ignore_errors is off); turning ignore_errors on skips such lines without listing themParse error giving the line and column, and the command fails; -R with fromjson? drops such lines silently
Listing the broken linesQuery log_level = 'invalid' in the same SQL viewRead with -f json,line to keep lines that are not JSON as plain-text eventsRead lines as text and test them with json_validRead lines with -R and use try fromjson catch
Before the first query on your own logWrite a JSON format file once, unless a built-in format (Bunyan, Pino, journald and others) matchesNone: JSON is recognized without a format file; a new option syntax and a small scripting language to learnNone; columns inferred from the first 20,480 recordsNone
Counting errors by endpointSQL GROUP BY on the format’s table, inside the viewer or with -n -c--freq path counts the values of a field; filter on the level firstSQL: SELECT … GROUP BY on read_ndjsongroup_by over the whole input, or paths piped to sort | uniq -c
Every record for one requestSQL WHERE, or step through one request with o and O once the format names its request-ID field--filter on the request-ID fieldSQL WHEREselect(.request_id == "…")
How you workFull-screen terminal viewer: timeline, jump to errors, filters, SQLCommand-line tool with level maps and summaries; no full-screen viewer documentedSQL shellCommand-line filter
Sources for triage and broken lines

Documentation checked October 20, 2026 (Kelora, November 11): lnav log formats, command-line options and release notes (0.10.0 and 0.11.2 for broken lines); DuckDB JSON loading and JSON functions; the jq manual and issue 884; the Kelora README, changelog and security policy. In an earlier hands-on check on October 3, 2026, before our research-only method, lnav 0.14.1 and angle-grinder 0.19.6 behaved as documented on one nine-line invented log, and both commands still finished with a success status. That record is history, not current testing.

Installation, maintenance and risk

Show 6 criteria
Installation, maintenance and risk
FeaturelnavKeloraDuckDB CLIjq
License and costFree; BSD-2-ClauseFree; MITFree; MITFree; MIT
Current release, checked November 11, 20260.14.1, September 5, 2026; 0.15.0 beta since September 28, 20262.2.1, October 8, 2026; commits to October 292.0.0, October 21, 20261.8.2, June 20, 2026
InstallStatic downloads for Linux, macOS and Windows; Homebrew, Snap, FreeBSD; distribution packages can be much olderDownloads for Windows, macOS (Intel and Apple silicon) and Linux; Homebrew tap, cargo; .deb and .rpmOne program for Linux, macOS and Windows; Homebrew or winget; Windows needs the Visual C++ RedistributableDownloads for Linux, macOS and Windows (Windows on ARM added in 1.8.2), Docker image, most package managers
Network behaviorREADME: sends nothing unless you ask; crash-log upload only on request; the 0.15 beta starts a local remote-control web server automaticallyNone: no networking or telemetry, enforced by a check in its build (its own statement)Installs and loads signed core extensions on demand by default (can be turned off)None documented
Published security recordNo published advisories; no security policyNo published advisories; security policy; fixes for a script-function flaw and zstd advisories are on the main branch but not in 2.2.1; no outside auditPolicy treats SQL and data files as trusted; engine advisories fixed in 1.1.0 and 1.4.2; AWS extension advisory, August 20261.8.2 fixed 16 CVEs, including one that let a malformed line containing a null byte pass (1.8.1 and earlier)
Maintenance and ownershipActive; mainly its creator, Tim Stack; crash past about 134 million lines fixed only in the 0.15 betaOne spare-time developer, best-effort support; AI agents write all of its code and its author does not review it; breaking changes in 2.1 and 2.2DuckDB Foundation holds the project; Amazon agreed in August 2026 to acquire DuckLabs, its developerjqlang organization; regular commits through October 2026 (checked October 20)
Sources for installation, maintenance and risk

Checked October 20, 2026, and again November 11, 2026 for releases, licenses and security pages: lnav README, releases, issue 1750 and security page; DuckDB release calendar, security policy, security settings, 2.0.0 release and Amazon announcement; jq 1.8.2 release notes and security advisories; Kelora releases, changelog, security policy and issues. A published record lists reported issues; it does not prove a tool is free of defects.

What matters when choosing

lnav is our Preferred Pick if you check the same application’s JSON-lines logs on your own computer again and again, and need to find the failing endpoint, count its errors and read every record for one request. It is a free, BSD-2-Clause terminal log viewer with statically linked downloads for Linux, macOS and Windows. Once lnav knows a log’s fields, you can jump from error to error, count errors by endpoint with an SQL query, and step through every message that shares a request ID. Its release notes say that a message lnav cannot parse is given an “invalid” level and highlighted, so a broken line stays in view and can be counted instead of quietly dropping out of your totals. Version 0.14.1, released September 5, 2026, was still the stable release on November 11. Log formats, release notes, 0.14.1 release

The lead is narrow. On five criteria we fixed before scoring, lnav scored 3.80 out of 5, the DuckDB command line 3.68, jq 3.63 and Kelora 3.60. lnav is ahead on what it lets you do once it is set up and on how it treats broken lines. It is behind on setup: lnav reads a JSON-lines log as records only when a format file describes it. It includes formats for some common JSON loggers, such as Bunyan, Pino and journald; for your own application’s fields you write a short JSON format file once and reuse it. Without one, lnav treats the file as plain text. Before you trust a total, count the invalid lines with a query, and remember that a line which does not start with { is shown as it is, without that flag. Log formats, release notes

For a one-off file whose fields you have not described, consider the DuckDB command line if you write SQL. It reads a JSON-lines file as a table with no format file, and by default a line that is not valid JSON makes the query fail rather than shrinking the count; its ignore_errors option skips such lines without listing them. Treat it as a powerful local engine: its security policy says it assumes data files come from a trusted writer and that corrupted files may crash it. Amazon agreed in August 2026 to acquire DuckLabs, the company that develops DuckDB; the project stays with the DuckDB Foundation under the MIT license. Version 2.0.0 was released on October 21, 2026. If jq is already on your computer, it does the same job with a short filter. A broken line makes jq report a parse error and fail. Use version 1.8.2 or later: it fixed a flaw that could let jq check only part of a malformed line containing a null byte. JSON loading, DuckDB security policy, Amazon announcement, DuckDB 2.0.0 release, jq manual, jq 1.8.2 release

Our Rising Star is Kelora, a newer command-line log processor, if you want to count and filter a JSON-lines log without writing a format file. Its first release was in May 2025. It is free and MIT-licensed, has downloads for Windows, macOS and Linux, and recognizes JSON lines on its own. Its documentation describes a filter on the level, --freq to count the values of a field such as the endpoint, and a filter on the request ID. A line it cannot parse is skipped with a message on standard error and the command still succeeds, and its -l option can leave such a line uncounted; --strict stops at the first bad line instead. It scored 3.60 against lnav’s 3.80. It is behind on maintenance: one developer works on it in spare time, its README says AI agents write all of its code and its author does not review it, and its minor releases have included breaking changes. It has no full-screen viewer, no SQL and no way to step through one request’s messages. Kelora README, changelog, security policy

angle-grinder, which scored 3.05, still makes a short counting pipeline, but its documentation says its JSON parser drops a line that is not valid JSON, and it has had no release since March 2025. Miller (3.25), VisiData (3.23) and sqly (3.08) also scored below the four tools above; the scores are in How we researched. angle-grinder README

These are researched judgments from documentation, release notes, source code, issue trackers and security advisories checked October 20, 2026, with releases and licenses read again on November 11, 2026, and security pages for lnav, DuckDB, jq and Kelora. In an earlier hands-on check on October 3, 2026, before we adopted our research-only method, lnav 0.14.1 and angle-grinder behaved as their documentation describes on one nine-line invented log: lnav showed a truncated line as an invalid row, and angle-grinder printed an error for it and left it out of the count. Both commands still finished with a success status, so neither works as an unattended check. That check is not the basis of the ranking. Earlier check record, October 3, 2026

Disclosure: Practical Picks articles are researched and written with AI, currently Anthropic’s Claude. lnav’s README thanks Anthropic for including lnav in its Claude for Open Source program, which gives qualifying open-source maintainers free Claude access, and Kelora’s README offers an agent skill for Claude Code. We have no relationship with the lnav or Kelora projects or Anthropic over this article, and those links carried no weight in our judgment. lnav README, Kelora README

Preferred Pick · repeated triage of one application’s logs

$0BSD-2-Clause · terminal viewer

Choose lnav when you check the same application’s JSON-lines logs again and again and want to move through them, count errors by endpoint and follow one request. Write its format file once, and count invalid lines before you trust a total.

Reasons to choose it

  • Broken JSON lines get an “invalid” level and stay in view
  • SQL counts, error-to-error jumps and request-by-request stepping in one tool
  • Static downloads for Linux, macOS and Windows

What to watch

  • Your own log’s fields need a format file; without one the log opens as plain text
  • Lines that do not start with { are not flagged
  • A crash past about 134 million lines is fixed only in the 0.15 beta

lnav is a terminal viewer for log files, maintained mainly by its creator, Tim Stack. It merges files by time into one view, builds an index of errors and warnings, filters messages, and exposes each recognized log format as an SQL table you can query from inside the viewer. The project offers statically linked downloads for Linux, macOS and Windows, and it is in Homebrew, Snap and FreeBSD packages; distribution packages can be much older, for example 0.11.2 in Ubuntu 24.04. It is BSD-2-Clause licensed. README and installation, releases, license

For JSON lines, lnav needs a format definition. Its documentation says that when a file matches no format it is treated as plain text, and that JSON-lines logs are handled by creating a format file that names the fields and how to display them. Built-in formats cover several JSON loggers, including Bunyan, Pino, journald, MongoDB, PostgreSQL’s JSON log and Rust’s tracing crate. In your own format file, name the request-ID field as the opid-field: lnav then lets you step forward and back through every message of that request with the o and O keys. You can keep the file in a folder of your choice and load it with -I, and run a query without the interface with -n and -c. Writing that file is the main cost of choosing lnav for an application whose log format it does not already know. Log formats, command-line options

Read the full lnav analysis and sources

How lnav treats a broken line is the main reason for our pick. Its release notes for 0.10.0 say a message that cannot be parsed is given an “invalid” log level with the faulty part highlighted, and the 0.11.2 notes say that a line in a JSON-lines file that does not start with { is shown as it is, without a parse error. So a truncated JSON record is flagged, while a stray plain-text line, such as a stack trace, is not. Query for log_level = 'invalid' and look for unparsed lines before you treat an error count as complete. In an earlier hands-on check on October 3, 2026, before we adopted our research-only method, lnav 0.14.1 showed a truncated line in a nine-line invented log as an invalid row with the position where parsing failed, and the command still finished with a success status. Do not use the exit status alone as a check. Release notes, earlier check record, October 3, 2026

lnav is actively developed. Version 0.14.1 came out on September 5, 2026, a 0.15.0 beta followed on September 28, and changes were still being committed on November 9. A September 12, 2026 report described a crash when a file passed about 134 million lines; the maintainer committed changes referencing it on September 14, and the 0.15.0 notes say indexing now stops at that limit with a warning. That fix is not yet in a stable release; on November 11 the report no longer appeared in the project’s list of open issues. The README says lnav will not send information to other systems unless you ask it to, and its crash-log upload command runs only when you call it. The 0.15.0 notes add that the beta starts lnav’s local remote-control web server automatically when the interface opens, on a port chosen by the operating system and with a random key, and records how to reach it in a file under ~/.lnav. Check that setting when 0.15.0 is released. The project publishes no security policy and has no published security advisories. Releases, issue 1750, release notes, README, security page

Choose Kelora, DuckDB or jq instead for a single file you will not see again, or if you would rather not learn a full-screen interface; lnav’s author has said that people are often tripped up when a file is not recognized as a log and is shown as text.

Used lnav? Share your experience to help other readers choose.

Kelora

Rising Star · a newcomer worth considering

$0MIT · command-line tool

Why Rising Star? First released in May 2025, Kelora scored highest among the newer tools that qualify: it reads JSON lines with no format file, counts the values of any field, lists one request’s records and reports lines it cannot parse. Its maintenance and its default of skipping broken lines keep it behind lnav, and every claim so far is its maker’s own.

Consider Kelora if you want to filter and count a JSON-lines log without writing a format file and you will read its error messages. It skips a broken line with a message and still exits with success, and its code is written by AI agents, so check its messages and count lines before you trust a total.

Reasons to choose it

  • Reads JSON lines with no format file
  • Counts the values of any field, such as the endpoint, and filters on the level or request ID
  • Downloads for Windows, macOS and Linux; no networking or telemetry by its own statement

What to watch

  • A broken line is skipped with a message and the command still succeeds; -l can leave it uncounted
  • One spare-time developer; its author does not review the AI-written code; minor releases have had breaking changes
  • Fixes for a script-function flaw are on the main branch but not in the current release

Kelora is a command-line log processor written in Rust by a single developer, Dirk Loss, and first released on May 24, 2025. It is MIT-licensed. Its README says it recognizes more than 20 log formats on its own, JSON among them, and splits each line into named fields. The project publishes downloads for Windows, macOS (Intel and Apple silicon) and Linux, and it is available through a Homebrew tap and cargo install. Version 2.2.1 came out on October 8, 2026, and its developer was still committing to the main branch on October 29. README and installation, releases, changelog, author named in the README of his earlier tool klp

The documented options cover the job without a format file. -j reads JSON lines, --filter 'e.level == "error"' keeps matching records, --freq path counts the values of a field, and a filter on your request-ID field lists one request’s records. Its README shows a level map and a comparison of message patterns before and after a time. There is no full-screen viewer, no SQL and no stepping through one request’s messages, and you learn a new option syntax and a small scripting language (Rhai) for anything beyond the options. README, security policy examples, changelog

Read the full Kelora analysis and sources

What it does with a broken line is documented, and it is less safe by default than DuckDB or jq. Its default is a resilient mode: a line that cannot be parsed is skipped with a message on standard error, and the command still exits with success unless no line parses at all. --strict stops at the first parse error. The 2.2.0 notes say that the -l level option can undercount parse errors on JSON and logfmt input; --strict turns off that shortcut. Reading with -f json,line keeps lines that are not JSON as plain-text events, where you can count them. Read the messages on standard error before you trust a total. Security policy, changelog, README

Kelora’s own account of its risks is plain. The README says AI agents write all of its code and tests, its author steers requirements without writing or reviewing code, and it is a single-developer spare-time project with best-effort support. Version 2.0.0 came out on June 18, 2026, and later minor releases have listed breaking changes. Its security policy says it has no network access and no telemetry, which a check in its build enforces, and that file writes are off unless you enable them. It also says limits on how much a script can compute are not implemented yet, and no outside audit has been done. The changelog for the next release lists two security fixes already on the main branch: a script function, parse_json(), evaluated its input as code, so a log line such as {"a": exit(7)} stopped the run and hid every later line, and the zstd library for .zst files had memory-safety advisories. The changelog says -f json was not affected. The project has no published security advisories, and its issue list showed no open issues on November 11. Security policy, changelog, README, issues

Choose lnav if you want to move through the log and follow one request; choose DuckDB if you write SQL; choose Kelora if you want counting and filtering without a format file and will read its error messages.

Used Kelora? Share your experience to help other readers choose.

DuckDB CLI

Also consider · a one-off file, if you write SQL

$0MIT · single program

Choose the DuckDB command line for a single JSON-lines file whose fields you have not described, if SQL comes easily to you. Leave ignore_errors off, so a broken line stops the query.

Reasons to choose it

  • Reads JSON lines as a table with no format file
  • A broken line makes the query fail by default
  • One program for Linux, macOS and Windows

What to watch

  • ignore_errors skips broken lines without listing them
  • Assumes data files come from a trusted writer; installs extensions on demand by default
  • Version 2.0.0 is new (October 21, 2026); Amazon agreed to acquire its developer

DuckDB is an analytical database engine, and its command line is a single program for Linux, macOS and Windows, available as a direct download, through Homebrew or through winget; on Windows it needs the Microsoft Visual C++ Redistributable. JSON support ships with most DuckDB builds and loads on first use. read_ndjson('app.jsonl') reads a JSON-lines file as a table, so one SELECT path, count(*) … WHERE level = 'error' GROUP BY path query finds the failing endpoint, and a WHERE request_id = … query lists one request. Column names and types are inferred from the first 20,480 records unless you set sample_size = -1. It is MIT-licensed. Version 2.0.0 was released October 21, 2026; 1.5.6 came out on September 28. Install, JSON overview, JSON loading, release calendar, 2.0.0 release

Its default on a broken line suits a count you intend to trust: ignore_errors is off, so a parse error stops the query instead of returning a smaller number. Turning it on skips bad lines without telling you how many, so do not use it for a count you will report. Its JSON reader documents no option for saving rejected lines, unlike its CSV reader, so to list the broken lines you read each line as text and test it with json_valid. There is no log view, timeline or error-to-error navigation; you work in SQL. JSON loading, JSON functions

Read the full DuckDB CLI analysis and sources

Treat it as a powerful local engine, not a sandbox. DuckDB’s security policy says SQL is executable code, that its file readers, JSON among them, “assume a well-formed file from a trusted writer”, that crafted or corrupted files may crash the process, and “Do not open data files from untrusted parties.” By default, DuckDB installs and loads a core extension when a query needs one; core extensions are checked against built-in signing keys, and SET autoinstall_known_extensions = false; and SET autoload_known_extensions = false; turn this off. Its published advisories include a 2024 file-access flaw fixed in 1.1.0 and an August 2026 advisory for the separate AWS extension, which this job does not use. Security policy, securing DuckDB, extensions, signing, security advisories, AWS extension advisory

Ownership of the developer changed in 2026. Amazon announced an agreement to acquire DuckLabs in late August 2026; Amazon says it is not acquiring the open-source project, which stays with the DuckDB Foundation under the MIT license. Techzine reported that the deal was set to take effect in early September 2026; we did not find a primary confirmation that it has closed. DuckDB’s August 2026 preview said 2.0 would bring a small set of breaking changes and a stable extension for serving databases over the network, and that it plans to store JSON values in a new format soon after 2.0. On November 11, the JSON loading page for the current documentation still listed ignore_errors and the 20,480-record sample, but check the 2.0 release notes before relying on other behavior above. Choose lnav if you will triage the same application’s logs repeatedly and want to move through them, not only query them. Amazon announcement, Techzine, August 26, 2026, 2.0 preview, August 17, 2026

Used DuckDB CLI? Share your experience to help other readers choose.

jq

Also consider · a one-off file, with jq already installed

$0MIT · command-line filter

Choose jq for a quick check with a tool you probably already have. Use 1.8.2 or later, and read its error message if you pipe its output into a count.

Reasons to choose it

  • No setup: reads one JSON value per line
  • A broken line gives a parse error with its line and column
  • Widely packaged, with release downloads for every major system

What to watch

  • Counting by a field takes group_by or a sort | uniq -c pipeline
  • fromjson? drops broken lines silently
  • Versions before 1.8.2 could pass a malformed line containing a null byte

jq is the widely used command-line JSON processor, maintained by the jqlang organization and released under the MIT license. Version 1.8.2, released June 20, 2026, is current; the project publishes binaries for Linux, macOS and Windows, including Windows on ARM since 1.8.2, and a Docker image, and most package managers carry it, sometimes in older versions. It reads one JSON value per line without any setup. select(.level == "error") filters records and select(.request_id == "req-42") shows one request. Counting by endpoint takes more: group_by over the whole input read with -s or -n and inputs, or passing the paths to sort | uniq -c. There is no viewer or navigation. Manual, 1.8.2 release, README

By default a line that is not valid JSON makes jq print a parse error giving the line and column, and the command fails. That protects a count, with one exception: in a shell pipeline such as jq … | sort | uniq -c, the count of the lines before the error still prints, and the pipeline reports success unless you set pipefail. Read the error message before you use the number. The manual documents -R, which reads each line as text, together with fromjson and try … catch. A common answer, jq -R 'fromjson?', silently drops lines that do not parse, which hides them from your count; try fromjson catch can print them instead. Manual, skipping non-JSON lines, issue 884, codeinthehole TIL, August 10, 2023

Read the full jq analysis and sources

jq’s 2026 security record is long, and nearly all of it is fixed in 1.8.2. That release lists 16 CVEs and other fixes, mostly memory-safety and denial-of-service problems in its parser and built-in functions. One bears directly on this job: in 1.8.1 and earlier, a null byte inside the input made jq check only the text before it, so a malformed line could pass as valid (CVE-2026-33948). If your system package is older than 1.8.2, use the project’s release. 1.8.2 release notes, security advisories, null-byte advisory

Choose lnav for repeated triage of one application, and DuckDB if counting and joining in SQL comes more easily to you than jq’s filter language.

Used jq? Share your experience to help other readers choose.

Which product are we missing?

Know a product that belongs in this article? Tell us what it does differently and why it deserves a closer look. Newcomers welcome.

Readers and product teams can suggest candidates. Suggestions inform our research; they do not guarantee inclusion or a recommendation.

Suggest a product

Opens an email draft with this article attached as a link. Nothing is sent until you send it.

No email app? Send the product name, website, and this article’s title to admin@practicalpicks.guide.

What people are saying

Public discussion is a check on project claims, not a measured reliability score. In a June 2024 Hacker News discussion of lnav, users praised its navigation, such as jumping to the next error or to the next message with the same request ID, and one shared a format file for Caddy’s JSON logs. Another said the interface overwhelmed them. lnav’s author replied that people are often tripped up when a file is not recognized as a log and is shown as text, which is what happens to a JSON-lines log without a matching format. Hacker News discussion, June 2024

In an April 2024 discussion of angle-grinder, one commenter described it as complementary to jq and most useful for live reports, another pointed out that jq can aggregate too, and others suggested lnav instead. jq’s issue tracker shows the long-standing question of how to get past lines that are not JSON: a 2015 request was answered with jq -R 'fromjson?', which skips them. Hacker News discussion, April 2024, jq issue 884

lnav’s September 2026 crash report came from the developer of a competing large-file viewer, who disclosed that in the report. Its details were disputed in the thread, and the maintainer then committed changes for the line limit. Kelora has little public discussion: two Hacker News launch posts, in December 2025 and June 2026, each drew one point, and we found no independent review. Our October 20 search for vulnerabilities, crashes, abandonment, ownership changes and disputes involving the tools found nothing beyond the items in this article that changed our conclusions, and a November 11 reread of release pages, and of the security pages of lnav, DuckDB, jq and Kelora, found only the Kelora items above; a search that finds nothing is not proof that nothing exists. No Practical Picks reader reviews have been received for these tools. lnav issue 1750

How we researched

This is public-source research. We did not install or run any tool for this review, open any real log, or measure speed or accuracy. No project supplied access, payment or input, and we use no affiliate links. We first compared lnav and angle-grinder on October 3, 2026, and rechecked every claim on October 20, 2026: releases, licenses, installation routes, documented behavior on broken lines, network behavior, security advisories, open reports and ownership. On November 11, 2026 we read again the release pages and licenses of every tool, the security pages of lnav, DuckDB, jq and Kelora, DuckDB’s 2.0.0 release and JSON loading page, and lnav’s open issues and commits, added Kelora, and read the documentation or project pages of Miller, VisiData, Nushell, sqly, Lazylogs, LogLens and LogSquirl. We read project documentation and release notes, GitHub repositories, release and commit feeds, issue trackers and security pages, one commit’s changes, the file loader of one tool we set aside, the OSV vulnerability database, Homebrew records, and Hacker News discussions.

We wrote this for a developer who has a JSON-lines log from their own application on their computer and needs to find the endpoint with the most errors, count them, read every record for one request and notice lines that are not valid JSON, without uploading the log. Before scoring, we fixed five criteria and their weights: fit for that triage job (30%), handling of broken lines (25%), setup and learning (20%), maintenance and continuity (15%) and local data handling and security (10%). lnav led with 3.80 out of 5, ahead of DuckDB at 3.68, jq at 3.63 and Kelora at 3.60; Miller scored 3.25, VisiData 3.23, sqly 3.08 and angle-grinder 3.05. The lead is narrow. It held when we gave broken-line handling more weight. When we gave setup more weight, as for a one-off file, DuckDB and jq moved ahead of lnav and Kelora drew level with it, which is why we name them for that use. With equal weights, lnav and jq tied.

This review added DuckDB and jq, which the October 3 draft did not cover. We scored three more established tools from their own pages and ranked them below the top four. Miller is a command-line record processor under the BSD-2-Clause license, available through apt, snap, Homebrew, MacPorts, winget, Chocolatey and Scoop. Its documentation says it reads JSON Lines, and its filter and count-distinct commands select records and count the values of a field. We found nothing in its documentation about what it does with a line that is not valid JSON, so we gave broken-line handling the low score we give any tool that does not document it. That score, and so Miller’s 3.25, could rise if a later check finds documented, visible behavior. VisiData is a terminal spreadsheet under the GPLv3 that opens JSON lines and draws a histogram of a column with one keystroke. When we read its loader code on October 20, a line that failed to parse after earlier rows had loaded became an error row. Its install page lists pip, Homebrew, conda and Linux packages; on Windows it points to the Windows Subsystem for Linux or MSYS2’s Python. It scored 3.23. sqly (MIT; Windows, macOS and Linux) loads a JSON-lines file into a one-column SQLite table, and you reach each field with json_extract(); its README does not say what happens to a broken line, and it scored 3.08. Miller file formats, Miller verbs, Miller README, VisiData install, sqly README

We did not score Nushell or ClickHouse local. The Nushell documentation page for from json shows an --objects option that reads one JSON value per line, but we found no page showing its commands for filtering and counting or what it does with a bad line. We could not read ClickHouse local’s documentation page. Nushell from json

We also set aside hl (a viewer that formats and filters JSON logs but does not count or group them), fx (an interactive JSON viewer that processes data with JavaScript, so counting by a field means writing code over the whole input), jless (a JSON viewer with no release since July 2023), Toolong (a terminal log viewer with no release since March 2024), gojq and jaq (reimplementations of jq’s language with the same strengths and limits) and log servers such as Grafana Loki, Graylog, VictoriaLogs and OpenObserve, which must be running and ingesting logs before you can query them and so do not suit a one-off check of a file.

We searched for newcomers on October 20 and again on November 11, 2026, in Hacker News launch posts, GitHub topic pages for log viewers and JSON lines, Rust crates and the related-tools lists of log projects. We read the project pages of these candidates: Kelora, Gonzo, Logdy, hl, klp, Lazylogs, LogLens, LogSquirl and sqly. Others that appeared only in launch posts and topic listings, such as lumtail, logsift and taffrail, described a different job or lacked Windows support, and we did not read their pages. We did not search app stores, Product Hunt or AlternativeTo, so a newer tool listed only there could be missing.

Kelora, a Rust command-line log processor, met all four of our conditions for a Rising Star: its first public release was on May 24, 2025, less than two years before this review; it can be downloaded now for Windows, macOS and Linux; its documentation and changelog show it doing the job; and we found no published High or Critical advisory, abandonment or false claim. Kelora is its author’s successor to klp, his earlier Python viewer, which he no longer maintains. It is a complete rewrite with its own name and releases. We found no independent coverage of Kelora. We did not score that as a weakness. Its maintenance score of 2.5 rests on its single developer, the unreviewed AI-written code and the breaking changes in minor releases. Gonzo, a terminal dashboard for live log streams from the observability company ControlTheory, first released in 2025, does not document counting by a field or listing broken lines, so it did not qualify; we scored it 3.03 for comparison. Logdy was first released in early 2024 and has had no release since June 2025. None of the other newcomers qualified. Lazylogs (first release March 1, 2026) filters by level, time and text but does not document counting by a field, and its page does not show Windows support. LogLens offers counting of query matches free, but its statistics command is Pro only, only its parsing engine is published as open source, and its page names the vendor only as Caelrith. LogSquirl is a desktop viewer that continues the older klogg project, adds a JSON table view only in its next release, and its macOS build needs Apple silicon. sqly’s first release date is not on its page, so it could not qualify as a newcomer, and at 3.08 it would not have outscored Kelora. Lazylogs, LogLens, LogSquirl

angle-grinder scored 3.05. Its README says its JSON parser drops a line that is not valid JSON. Its latest release, 0.19.6, came out on March 29, 2025; the download reports itself as 0.19.5 because the release was built from a commit that still carried the old number, a labeling slip and not a different build. It has one maintainer, no Windows download, no published advisories, and a July 2024 report that a multi-field assignment fails silently is still open. README, releases, version commit, issue 208

On October 3, 2026, before we adopted our research-only method, we ran a small local check: official Linux downloads of lnav 0.14.1 and angle-grinder 0.19.6 against one invented nine-line log with one truncated line, plus a screen of a newer browser-based log viewer. That record is kept as dated history and is linked where it is cited. It supports the documented behavior described above and does not decide the ranking.

ToolPublic sourcesWhat remains uncertain
lnavFormat and command-line documentation, release notes, releases and commit feed, README, license, issue 1750, security page, Ubuntu package records.Final 0.15.0 behavior, including the remote-control server; how often unparsed plain-text lines go unnoticed; behavior on very large files.
KeloraREADME, changelog, security policy, releases, commit feed, issue list, Hacker News search.How it treats a truncated line in practice; the next release’s security fixes; behavior on very large files; whether a lone developer will keep it going.
DuckDB CLIJSON overview, loading and function documentation, install page, release calendar, 2.0.0 release page, security policy and advisories, extension and security settings, Amazon announcement, Techzine report, 2.0 preview.Changes in 2.0.0 beyond its JSON loading page; the error text for a broken JSON line; whether the DuckLabs acquisition has closed.
jqManual, 1.8.2 release notes, security advisories, README, issue 884, a third-party write-up of its parse error.Which jq version your system package ships; exact behavior after a parse error in each version.
angle-grinderREADME, license, releases, commit feed, version commit, issue 208, security page.Whether development will continue; behavior on large streams.
MillerFile-format and verb documentation, README, license.What it does with a line that is not valid JSON; behavior on very large files.
VisiDataHome and install pages, October reading of its JSON loader code.Whether selecting rows by a field value suits this job; behavior on a broken line in practice; Windows use without WSL.
sqlyREADME, license, commit feed.What it does with a line that is not valid JSON; when it was first released.

None of this establishes speed, memory use or accuracy on your logs. Next substantive review is due December 19, 2026, or sooner after lnav 0.15.0 or Kelora’s next release. Found a changed release or behavior? Contact the editors.

Sources and evidence

Sources were checked October 20, 2026 unless another date is given; Kelora and the November 11 rechecks are dated. Project material establishes documented features and published releases, not behavior in your environment. Community posts are attributed opinions, not representative samples. The October 3, 2026 check record is a historical observation, not current testing.

No screenshots are shown. This comparison rests on public documentation and records; an earlier check on October 3, 2026, before our research-only method, used one invented log. Earlier check record, October 3, 2026.