AI debugging agents
A failing flow, a silent 500, a setup or config issue, tell Upliftr the way you'd brief a teammate. A team of AI agents investigates: they reproduce it in a real browser, or probe your services and logs when it's infra, and trace it to the exact cause, the UI, the API, the backend line, or the broken dependency, then file a deduped ticket with the fix. They only ever read, never touching your code, data, or infrastructure. Self-hosted, your keys.
Source-available · BYO Anthropic key · cloud or self-hosted
The autonomous agent
Give an agent a symptom, or just “find what's broken,” and it investigates by itself: it fans out across all your log sources, probes your services, and follows the failure from the UI to the failing request to the exact backend line, then reports the cause and the fix. No recipe, no map, it reasons about whatever environment you drop it on, and it's read-only by design, so you can run it against production.
What it does
You chat with agents that work across your flow, your servers, and your infra. They drive a real browser to reproduce a bug and root-cause it to the exact backend line, or probe your infrastructure directly when the problem isn't the UI: a firewall, a down service, a DNS or TLS issue. Then they file the fix. Read-only, they never change a thing.
Tell them what's broken or what to check, in plain English. Or fire an autonomous investigation: drop the agent on a symptom and it root-causes on its own, across logs, infra, and the UI, following the trail until it has the cause. Read-only, always.
The autonomous agent →A failing request's trace_id is correlated to your server logs: the backend error, the likely cause, a suggested fix. Past the symptom, to the line that threw.
Point it at Loki, Datadog, Elasticsearch, an HTTP endpoint, or a plain log file. Upliftr stitches the trace itself, no instrumentation project, no spans to wire up first.
Connect logs →The agents drive a real browser, headless on Linux, macOS, or Windows, re-grounding against the live DOM each run, so they self-heal instead of breaking on a stale locator. A confirmed reproduction, not a guess from telemetry.
How it works →Confirmed failures become root-caused tickets on the built-in board, and optionally GitHub, GitLab, or Jira. Re-runs comment, never duplicate.
Filing issues →Anything the agents verify they save as a plain-English check they re-run forever, in your IDE (MCP for Cursor & Claude Code) and as a GitHub Action, GitLab CI, and CLI that gate every PR/MR.
In CI & your IDE →Agents, not dashboards
Sentry and Datadog show you an error after it happened, from telemetry you had to wire up, and each stops at its own layer. Upliftr's agents do the opposite: they drive your real app, reproduce the failure on demand, and follow it across the UI, the API, and the exact backend line, then hand back the cause and the fix. Active, not passive, and read-only, so they never change your systems.
UI: what the user hit
“Save profile” returned a silent 500. Nothing shown to the user.
Backend: the real cause
NullPointerException in upload.py:142. Avatar saved before the user row commits.
Filed: GitHub #284
Cause + suggested fix, deduped. The ticket was waiting for you.
What it remembers
When the agents confirm a flow, they keep it as plain-English YAML you can read and edit, no selectors, no page objects. Cases run in order and share fixtures, so “sign up, then log in as that user” just works, every time they re-check.
Your infra. Your keys. Your data.
Every other AI tool like this is closed SaaS. Your app and data go to them. Upliftr is source-available and runs entirely in your own environment with one command. Bring your own Anthropic key; nothing leaves your network.
One command, curl -fsSL https://get.upliftr.io | bash, pulls prebuilt images and brings up the whole platform behind trusted HTTPS. Re-run to upgrade.
Your Anthropic key drives the agents. No per-seat AI markup; you pay your provider directly.
The agents observe and diagnose, they never mutate your code, data, or infrastructure. Per-org isolation, restricted egress, read-only log access, encrypted secrets.
Built in
An MCP server puts the agents inside Cursor and Claude Code, debug a flow without leaving your IDE.
A GitHub Action / GitLab CI and a CLI gate every PR/MR with a JUnit report and a meaningful exit code.
The built-in board, plus GitHub, GitLab, or Jira, deduped by fingerprint.
Read-only probes debug reachability, DNS, TLS, and dependency failures, not just UI bugs. Server-log correlation names the backend line.
Pricing
Source-available: run the whole platform yourself, bring your own key.
For any noncommercial use. Everything: chat debugging, backend root-cause, saved checks, CI gating. Bring your own LLM key.
Get started →For business & production use. A commercial license with terms that fit your team, plus deployment guidance.
Deploy it →Describe the bug in plain English. A root-caused fix out, automatically. Self-hosted, bring your own key.