ai log analyzer
Read the envelope before you send it
Logier is a desktop AI log analyzer for macOS and Windows. Open a log file in any format — including one nobody has written a parser for — and get a plain-English explanation of what broke, when, and why. No agents, no ingestion pipeline, and no Logier inference service.
Bring your own key · or point it at a model on your own machine
what you can see
14:02:07 error webhook retry exhausted (job wh_3ab1) 14:02:09 error webhook retry exhausted (job wh_3ab2) 14:02:11 warn pool acquire 30011ms
what it says
The retry queue filled after the database connection pool was exhausted three minutes earlier. The errors you are looking at are downstream noise, not the cause.
Searching returns lines. Analysis returns a claim.
A conventional log tool searches. You supply a keyword, it returns matching lines, and the work of understanding what those lines mean is entirely yours. That is fine when you already know what you are looking for, and close to useless when you don’t.
An AI log analyzer interprets instead. Given a failure, it reads the entries around it — what ran in the seconds before, which lines share a request or job ID, whether this signature has appeared in the file before — and produces an explanation rather than a selection.
Asked about a burst of webhook timeouts, it does not reply “the webhook timed out.” It reports that the connection pool saturated ninety seconds earlier, that every failing job targets the same downstream host, and that the first failure follows a deploy marker by nine seconds. That is a hypothesis you can test — which is a different kind of object from a list of matching lines.
The envelope, opened.
Every tool in this category says your logs are safe and none of them shows you the request. Here is the whole payload: three blocks, in this order, with the actual size bound on each. The app computes these same numbers and shows them to you before anything is sent.
- 01
[LOG WORKSPACE METADATA]What the file is
The filename, which parser claimed the file, how many structured lines were parsed, and — if you narrowed the range — the time window you chose. No log content.
bound — Four lines
- 02
[HIGH-PRIORITY MULTI-LINE ANOMALIES FOUND]The lines that look like trouble
A scan for lines matching failed, error, exception, denied, permitted, critical or socket — case-insensitive — each captured as a four-line block: the two lines before it, the line itself, and the one after, so a stack trace arrives with its head attached.
bound — At most 40 excerpts
- 03
[RAW DATA INITIAL PREVIEW]The top of the file
A straight prefix of the file, or of the time window you selected. This is the block that makes the answer contextual rather than a reading of forty disconnected errors — and it is the block most tools do not admit to.
bound — Half your token budget, 120,000 characters maximum
For a file smaller than the preview budget, block three is the whole file. Logier says so plainly — the context indicator reads “full file” rather than implying a redaction that isn’t happening. Above the budget it shows how much of the file is covered and what share that is, and a time-range slider moves the window onto the minutes you care about.
None of this applies at all when the provider is set to Local. The excerpt goes to an endpoint on your own machine, and the question of what a vendor retains stops being a question.
Four places the excerpt can go.
You connect the provider you already trust, with a key you already have. There is no Logier inference service in the middle, and keys are stored locally and excluded from settings sync.
- ChatGPT
- OpenAI, with your own API key.
- hosted
- Claude
- Anthropic, with your own API key.
- hosted
- Gemini
- Google, with your own API key.
- hosted
- Local
- Any OpenAI-compatible endpoint on your own machine — Ollama, LM Studio, or anything else that speaks the protocol. Nothing leaves the machine.
- on-device
Most of Logier never asks a model anything.
The AI layer sits on top of a tool that already works without it. Everything here runs before you connect anything, and keeps working if you never do.
- 01
Automatic format detection
Eleven named parsers plus a heuristic fallback for formats nobody has written a parser for. No model involved, and no configuration.
- 02
Clustering
Thousands of identical errors collapse into one entry with a count, so the shape of the file survives contact with a retry loop.
- 03
Timeline
Entries bucketed by minute and plotted per level, so a spike is visible before anyone has to describe it. A minute with no entries is drawn as a break, not a zero.
- 04
Anomaly detection
Frequency spikes, new signatures, silent gaps and outlier values — all trailing-window statistics computed locally, with no model and no network call.
Four categories wearing one name.
“Log analysis tool” covers four fairly different things solving fairly different problems. Most teams end up using more than one, and the mistake is reaching for a platform when you have one file and one question.
| Category | Setup | Your log data | Best for |
|---|---|---|---|
Desktop AI log analyzers Logier | Open a file | Stays on your machine | Debugging a specific failure, reading a log someone handed you, working with sensitive or proprietary logs |
Cloud observability platforms Datadog, Splunk, New Relic | Agents, ingestion pipelines, indexing | Uploaded and retained | Continuous production monitoring, alerting, long-term retention across a fleet |
Self-managed log stacks ELK, Grafana Loki, Graylog | Infrastructure you run and maintain | Your servers | Centralised search across many services when you want to own the stack |
Terminal log viewers lnav, klogg, grep | None | Stays on your machine | Fast local search and tailing when you already know what you're looking for |
the honest limit
It answers about a file, not about your fleet.
The model sees one file and the window you gave it. It cannot correlate across services, it has no history of last Tuesday, and it will not page anyone at three in the morning. If the question is “is production healthy”, this is the wrong tool and a platform is the right one.
It is also worth saying plainly that a language model can be confidently wrong. Logier’s answer is a hypothesis with the lines it was drawn from attached, which is why the findings panel and the timeline are computed by local statistics rather than generated — the numbers you check the answer against never come from the model.
faq
Common questions
keep reading
Go deeper
- 01
AI root cause analysis
The feature page: what Copilot refuses to answer, and why the refusals are the interesting part.
- 02
Unknown log format analyzer
How a log from proprietary software becomes parsed entries the model can reason over.
- 03
Anomaly detection
The findings the model reasons over — statistics, computed locally, with no model involved.
- 04
Local-first log privacy
Every outbound request the app can make, listed. Including this one.
Ask your logs what went wrong.
On macOS and Windows, with your own API key — or a model that never leaves your machine.