Log Reducer

Your AI coding agent is spending thousands of tokens reading raw logs — DEBUG spam, health checks, duplicate lines, framework stack frames, UUIDs. Those tokens are gone for the rest of the session. The agent has less room to think, generates worse code, and hits its context limit faster.

Log Reducer sits between the log and the AI. It reduces the file down to just the signal — errors, warnings, state changes, unique events — typically cutting 70-90% of tokens. The raw log never enters the AI's context.

It runs as an MCP server (the AI calls reduce_log with a file path) or as a CLI (pipe any log through it). No API keys, no network calls — deterministic text transforms that run instantly.

Example

You're running your FastAPI dev server. You click around, hit a 500 error, and copy the terminal output into a file. It's 218 lines — mostly a wall of framework stack traces:

218 lines, 1185 tokens  →  51 lines, 310 tokens  (74% reduction)

Here's what the tool does to the stack trace. This is a real Python exception group with uvicorn, starlette, and FastAPI frames:

Before — 95 lines of stack trace, full C:\Users\...\.venv\Lib\site-packages\ paths:

    |   File "C:\Users\imank\projects\video-editor\src\backend\.venv\Lib\site-packages\
             uvicorn\protocols\http\httptools_impl.py", line 426, in run_asgi
    |     result = await app(  # type: ignore[func-returns-value]
    |              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "C:\Users\imank\projects\video-editor\src\backend\.venv\Lib\site-packages\
             uvicorn\middleware\proxy_headers.py", line 84, in __call__
    |     return await self.app(scope, receive, send)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    ... 85 more framework lines ...
    |   File "C:\Users\imank\projects\video-editor\src\backend\app\routers\exports.py",
             line 745, in list_unacknowledged_exports

After — your code preserved, framework collapsed, duplicate traceback gone:

    |   [... 10 framework frames (uvicorn, fastapi, starlette, contextlib) omitted ...]
    |   File "app/middleware/db_sync.py", line 107, in dispatch
    |     response = await call_next(request)
    |   [... 6 framework frames (starlette, contextlib) omitted ...]
    |   File "app/main.py", line 97, in dispatch
    |     response = await call_next(request)
    |   [... 16 framework frames (starlette, fastapi) omitted ...]
    |   File "app/routers/exports.py", line 745, in list_unacknowledged_exports
    |     exports=[
    |   File "app/routers/exports.py", line 746, in <listcomp>
    |     ExportJobResponse(
    | pydantic_core._pydantic_core.ValidationError: 1 validation error for ExportJobResponse
    | project_id
    |   Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType]

Traceback (most recent call last):
  [... duplicate traceback omitted ...]

The bug is clear: exports.py:745 passes project_id=None to a Pydantic model that expects an int. Three framework frames, not 95. No C:\Users\...\.venv\ paths.

(Full before/after | How the funnel pattern works for larger logs)

Setup

Step 1 — Install

npm install -g logreducer

Step 2 — Add MCP server to your project

Run this in your project root:

claude mcp add logreducer -s project -- npx -y logreducer --mcp

This writes the config to .mcp.json (the file Claude Code reads). Or add it manually:

{
  "mcpServers": {
    "logreducer": {
      "command": "npx",
      "args": ["-y", "logreducer", "--mcp"]
    }
  }
}

Step 3 — Add AI instructions

Tell Claude Code: "Follow the integration guide at https://github.com/launch-it-labs/log-reducer/blob/master/docs/agent-integration.md" — it will add the right instructions to your CLAUDE.md and set up the /logdump slash command.

Verify it worked: "What MCP tools do you have?" — it should list reduce_log.

That's it. Your AI agent now reduces logs automatically instead of reading them raw.

How to use

Once set up, you don't need to learn any commands or parameters — the AI handles everything automatically. There are just two things to know:

Sharing logs with the AI

Copy a log to your clipboard, then type /logdump in the chat. The raw log is saved to a temp file and reduced automatically — it never enters the AI's context. This is the recommended way to share logs.

You can also point the AI at a file: "check the errors in /var/log/app.log" — it will call reduce_log on it instead of reading it raw.

CLI (for scripts and piping)

You can also use it directly from the command line, outside of an AI session:

logreducer < app.log > reduced.log
kubectl logs my-pod | logreducer
logreducer --level error --context 10 < app.log

How it works

Everything below is for the curious — you don't need any of this to use Log Reducer.

What it does to your logs

Biggest impact first:

• Noise filtered — health checks, heartbeats, progress bars removed (DEBUG/TRACE lines kept — the AI chooses when to exclude them via level filter)
• Stack traces folded — 80 frames → your code frames + [... N framework frames omitted ...]
• Repeated lines collapsed — 6 similar lines → one template with varying values listed
• Log prefixes factored — 8 lines sharing timestamp - module - LEVEL → 1 header + indented messages
• Repeating blocks detected — 5 identical 3-line blocks → 1 block + count
• IDs shortened — UUIDs, hex strings, JWTs, tokens → $1, $2, ...
• Timestamps simplified — 2024-01-15T14:32:01.123Z → 14:32:01
• Test output collapsed — runs of PASS lines → count summary; FAIL lines always preserved
• Domain-specific — pip installs, Docker layers, HTTP access logs, retry blocks, log envelopes each have dedicated collapsers

19 transforms, applied in sequence. Rule-based, deterministic, no API calls required. One dependency (@modelcontextprotocol/sdk). Optional query param uses Claude for targeted extraction (requires ANTHROPIC_API_KEY).

[x7] [CacheWarming] Warmed tail of large video ({N}MB) | N = 2574, 3139, 2897, 3063, 2490, 2996, 3043

Multi-turn investigation

The tool isn't just a one-shot reducer. It supports a funnel pattern that lets the AI investigate a large log file in multiple targeted passes — spending ~1,000 tokens total instead of 5,000+ from a blind dump. The AI does this automatically, but here's what's happening under the hood:

Step 1: SURVEY → reduce_log({ file, tail: 2000 })           ~50 tokens
  If the reduced output exceeds the threshold (default: 1000 tokens),
  the tool automatically returns an enhanced summary instead of the full
  output: unique errors/warnings with counts, time span, and components.

Step 2: SCAN   → level: "error", limit: 3                   ~200 tokens
  See first 3 errors with context. Note timestamps.

Step 3: ZOOM   → time_range: "13:02:28-13:02:35", before: 50  ~500 tokens
  50 lines leading up to the first error — the causal chain.

Step 4: TRACE  → grep: "pool|conn", time_range: "13:00-13:05",  ~300 tokens
                  limit: 15, context: 0
  Follow the connection pool thread.

Total: ~1,050 tokens. The agent found the root cause (connection pool exhaustion from a batch job) without ever loading the full log.

See docs/agent-integration.md for the full parameter reference and filter details.

Design decisions

• File-path workflow — the MCP tool accepts file paths so raw logs never enter the AI's context. Only reduced output crosses into the conversation.
• Token reduction over line reduction — stats reported in tokens, not lines, since that's what matters for AI context windows.
• Generality over coverage — new transforms are scored by how broadly they apply. A pattern that only helps one application's logs gets flagged as bias risk and skipped, even if it would improve that specific case.
• Transform order matters — IDs and timestamps are shortened before dedup so lines differing only by those values become identical. Noise is filtered before prefix factoring so separator lines don't break grouping.
• Each transform is independent — pure function in, string out. Easy to add, test, and reorder without touching the rest of the pipeline.
• Minimal dependencies — pure TypeScript, one runtime dependency (@modelcontextprotocol/sdk).

Contributing