Capture AI coding traces where the work happens.

FML's local plugin records prompts, tool calls, file edits, setup, cost, and outcomes across Claude Code, Codex, and Gemini. Those traces become reusable team context before they become dashboards.

Get started

Local trace store

Session data captured on device

status

local first

Claude Codecaptured

Codexcaptured

Geminicaptured

trace record

checkout retry annotation

accepted

prompt + replies

tool calls

files touched

cost + tokens

MCP query

Ask for prior sessions and task context from the local store.

Injected context

Programmatically attach compact context before the next agent run.

Trace export

Send customer-owned records to eval, annotation, or audit pipelines.

What the plugin captures

A trace is the structured record of what the engineer asked, what the agent tried, what changed, and whether it worked.

Prompts and replies

The user request, assistant plan, follow-up turns, and final outcome stay attached to the session.

Tool calls and results

Shell commands, file reads, searches, edits, MCP calls, and failures become a replayable timeline.

Code and repo context

Files touched, branches, diffs, related commits, PRs, reviews, and reverts explain what the agent changed.

Cost and model telemetry

Token counts, model mix, subagent use, context pressure, and expensive loops are captured per session.

Setup and workflow signals

Skills, hooks, MCP servers, rules, permissions, and commands show which local setups produce better runs.

Outcome labels

Accepted work, rework, stuck sessions, review defects, and repeated failures become retrieval and eval signals.

The same record powers context, measurement, and export.

FML does not stop at observing sessions. The trace is linked to repo history and outcomes, then used to improve the next run.

Capture locally

FML Observe runs beside Claude Code, Codex, and Gemini. Engineers keep their normal tools while traces land in a local SQLite store.

Sync when useful

Teams can sync traces to FML for org search, session review, setup comparison, Slack answers, MCP access, and integration linking.

Inject back into agents

Prior sessions, repo history, rejected approaches, tests, docs, and review constraints are packaged into compact task context.

Export for pipelines

Enterprise teams can route customer-owned traces to S3 or internal pipelines for evals, audit, annotation review, and training workflows.

Keep the full trajectory, not just the label.

Teams using Claude Code for annotation or environment work can capture real task trajectories without forcing annotators into a new surface. The customer owns the traces and can use them for evals, audit, or training pipelines.

Capture annotation work happening inside Claude Code instances

Preserve full task trajectories instead of only final labels

Compare prompt, tool, and setup patterns across annotators

Export trace records to customer-owned eval or training pipelines

Export traces to S3 or customer pipelines in a standard trace schema, including ATIF-style workflows when a team needs interoperability.

Start with local capture. Compound into team context.

Get started