Local-first macOS (primary) No account · no hosted traces

Fork the step that broke. Prove the fix — on your machine.

LogMotif is a local agent lab: record instrumented model and tool calls, fork from a step, replay tools from tape (VCR), compare branches, and get a prove-the-fix verdict — without uploading runs to a vendor cloud.

Request early access See the loop

Desktop app · Python attach (best path) · TypeScript attach · mock mode offline

LogMotif

listening 127.0.0.1:7741 · 1 workflow · 3 runs

Attach New Run

Timeline Fork & Compare

Checkpoint · run start #1

web_search · researcher #2

ModelCall · researcher #3

ModelCall · critic #4

ModelCall · writer #5

Checkpoint · completed #6

Original · main

Critic: “Sources are thin. Confidence is low on the third claim; needs verification.”

Fork · fix-critic

Critic: “Be stricter about factual accuracy. Flag any claim without a primary source.”

✓ Fix verified · content divergence · first meaningful change at critic

Playing: record → fork → prove

Demo

The debugging loop

Record → pick a bad step → fork with a tweak → compare → verdict. The interactive mock above illustrates that path. A real screen recording will replace this section when we publish one.

Product walkthrough coming soon

Until then, use the mock in the hero — or request early access and try the desktop app yourself.

Illustrative mock · not a live recording of the app

How it works

Turn a failed run into a reproducible experiment

Orchestrating agents is getting easier. Understanding why a run failed still isn’t. LogMotif is a local flight recorder for that loop.

Record

Instrumented model calls, tool calls, and state updates land on an append-only timeline — from in-app workflows or an attached agent that emits events.

Fork

Select a step, fork with an instruction tweak or model override. The event prefix is copied; remaining steps re-run on a new branch.

Replay tools from tape

On fork, recorded tool results can replay from the VCR tape so you don’t re-hit search/HTTP APIs. LLM steps after the fork still run live.

Prove

Side-by-side compare with local text similarity (noise vs content). Verdicts: improved, regressed, unchanged, or inconclusive — not magic.

Features

What ships in the desktop lab today

⏱

Event-sourced timeline

Live streaming events, inspectable payloads, checkpoints, estimated cost per span, and a tape player to scrub the run.

⑂

Fork & compare

Branch from a selected event. Diff original vs fork. Structural, content, and noise classifications — local similarity, not embeddings.

✓

Prove-the-fix verdict

After compare, a heuristic verdict: improved, regressed, unchanged, or inconclusive — a signal to iterate, not a formal proof.

◎

Compose or attach

Author multi-step workflows in-app, or stream events from your own agent. Python attach includes live fork re-execute; TypeScript is record-focused today.

⌀

Keys stay local

OpenAI, Anthropic, Ollama, and Tavily can route through the app. Secrets in the system keychain. Mock mode needs no API keys.

⇄

Export & share

JSON run export/import and self-contained HTML tape files for demos and bug reports — no LogMotif account required.

Attach SDK

Your agent. Our debugger.

Keep writing agents in Python (or TypeScript). Stream newline-delimited JSON to LogMotif on 127.0.0.1:7741. Fail-safe by default — if the app isn’t running, your agent still runs (nothing is recorded). Frameworks like LangGraph work when you wrap nodes with spans; we don’t silently instrument whole ecosystems.

Request early access

python · attach demo localhost:7741

import agentweave  # package name (product is LogMotif)

@agentweave.record(name="my-agent", topic="Research")
def run(prompt: str):
    with agentweave.span("researcher"):
        agentweave.log_llm(
            prompt=prompt,
            content="…",
            model="gpt-4o-mini",
        )
    return "done"

# With LogMotif open: run appears · fork · compare

Why local

Debug-first. Data stays on your machine.

Hosted tracers are great for teams shipping agents at scale. LogMotif is for builders who want the experiment loop — compose, run, fork, prove — without a SaaS account for the debugger itself.

Capability	Typical hosted tracer	LogMotif (today)
Where run data lives	Vendor cloud	Your disk (SQLite)
Debugger account	Usually yes	No
Fork + re-execute	Varies by product	In-app + Python attach
Tool VCR on fork	Sometimes	Yes (recorded tools)
In-app workflow composer	Uncommon	Yes
Prove-the-fix verdict	Uncommon	Heuristic, built in

Note: live LLM/tool calls still use your providers (or mock mode / Ollama). “Local” means the debugger and traces — not that the internet is optional for every model.

Early access

Private beta — macOS first

No public download yet. We’re onboarding builders who will actually attach real agents. Mock mode works with zero API keys. Windows and Linux are possible via Tauri but not the focus of early access.

Email for early access See the product loop

hello@logmotif.com · logmotif.com · traces stay on your machine