Know what your AI spend delivered
The dashboards show users, models, and volume, but not whether the spend worked. FML follows spend along the path it took — intent, code, PR, deploy, outcome — and flags the expensive sessions with nothing linked.
Customer A
$850 / engineer / month
A large team with a standard AI coding rollout.
Customer B
$5,200 / engineer / month
A smaller, agent-heavy engineering team.
Different workloads may explain the gap. Usage alone cannot distinguish productive work from repeated discovery.
Where the spend went
Spend tied to the path it took, with the expensive sessions that produced nothing surfaced the week they happen.
Where the spend went · this week (illustrative)
$10.0k
Render-queue rewrite
$4.2k
shippedCheckout retries
$1.1k
mergedSearch reindex
$3.8k
stalled · no linked outcomeOnboarding flow
$0.9k
shipped$3.8k with no linked outcome, surfaced the same week it happened.
Every session and dollar in one place
Spend, tokens, model mix, and session activity — per repo, per engineer.
Activity
Last 30 days
Team briefing
Your team ran 128 sessions across 9 active repos. Checkout has the richest setup, 7 skills were used in sessions, and estimated spend is $3.8k.
Recent sessions
Findings
Large-context sessions
6 sessions exceeded 150k tokens before the first useful edit.
Model & token mix
Top models · 8.5M tokens
Practice library
Skills, commands, hooks, MCP
Connected systems
Signals attached to sessions
Talk to FML from
Slack, MCP, and CLI
Ask what the team shipped, what a session cost, or why a change landed. FML answers from the work record, posts a morning digest, and sends alerts from connected sources like Sentry and Linear.
2 sessions touched retry logic yesterday · view sessions
What FML found in 600 sessions
~32%
of this team's AI spend went to repeated discovery and known dead ends
600
sessions
3
engineers
30 days
captured
Repeated discovery means the model would have acted differently with information already in the team's history, as opposed to a logic failure. Across customers and models the avoidable share runs up to ~40%; the share recovered at install time, before any tuning, is ~25–30%.
This is a retrospective classification, not realized savings. The pilot measures the before-and-after directly: a baseline with injection off, then injection on, compared on cost per shipped outcome with an agreed quality guardrail.
How it deploys
A plugin and CLI on engineer machines capture sessions; delivery systems connect for outcomes.
Plugin and CLI capture
Installs on engineer machines, captures sessions from the coding agents your team already runs, and syncs them for analysis. Git, CI, and review stay as they are.
Gateway capture
Capture at the LLM gateway layer — LiteLLM, OpenRouter, Cloudflare AI Gateway — for fleets of headless agents.
The FML offering
Memory and attribution are live today. Router evaluation and historical evals are coming soon, built on the same install.
Engineering memory
Agents inherit the team’s history before they generate the first token: prior work, failed approaches, decisions, production impact.
Engineering attribution
Spend tied to the path it took: intent, routing, code, PR, deployment, outcome, cost.
Router evaluation
A router picks a model. FML checks whether that pick produced a better engineering outcome, replayed against your own sessions and PRs.
Historical evals
Recurring tasks and past failures become evals, so agent changes are tested against your real work instead of benchmarks.
Security and deployment
FML cloud
The fastest path to a pilot.
Single tenant
A dedicated environment for your organization.
Your cloud
Capture, storage, and context inside your account.
What's captured?
Session transcripts, tool calls, and diffs from registered harnesses.
What's excluded?
Secrets and credentials, redacted at capture.
Who sees raw sessions?
The author, plus roles your organization designates.
How is context scoped?
By project and repository. Injection follows the repos an engineer already works in.
Can engineers inspect it?
Every injected fact is source-linked and reviewable by the engineer it's shown to.
Retention & deletion
You own the raw data. Windows are configurable; deletion is in your control.
Compliance
SOC 2 program run on Vanta, with a live trust center showing current status and controls.
The pilot
Measure the baseline, turn context on, and compare cost per shipped outcome with an agreed quality guardrail.
STEP 1
Capture the work
Measure real spend and shipped output with injection off.
STEP 2
Turn context on
Measure the same team and the same outcome again.
STEP 3
Expand or stop
Compare against the threshold agreed at the start.
Install FML on a few machines with injection off. Thirty days later you have a real baseline, we turn injection on, and we compare.