Skip to content
NewApproval policies for MCP tools

Agent operations

Every agent run, traced.
Every risky action, approved.

Fernline is the control room for AI agents in production: see every step and token, route risky tool calls to a human, and keep LLM spend inside budget.

SOC 2 Type II · GDPR · EU or US data residency

Example Fernline console: runs, success rate, cost by model, the approval queue, a run trace, budget, failing agents, alerts and live activity.

Runs

86

1.1% vs the day before

Success rate

96.6%

2.5% vs the day before

Runs over time

Last 14 days

  • Succeeded
  • Needed approval
  • Failed

Cost by model

Last 30 days

30 days$594.17
  • Claude Opus 5.5$343.8158%
  • Claude Sonnet 5.5$144.0124%
  • GPT-5$64.2911%
  • Claude Haiku 5.5$37.016%
  • GPT-5 mini$3.091%
  • Gemini 2.5 Flash$1.960%

p95 duration

1m 43s

70% vs the day before

Spend by agent

Month to date

Approvals 4 pendingExample data
  • Issue refund $420 to customer #8812

    Refund Assistant

  • Publish status update for incident #331

    Incident Summarizer

  • Run SQL DELETE on staging.sessions_2024

    Data Cleanup Bot

  • Send email to 402 contacts

    SDR Outreach

Latest trace

Refund Assistant

Refund Assistant
Plan
lookup_order
refund-policy.md
Decide refund
Human approval: Refunds over $250
AgentLLM callTool callRetrievalHuman approval
Agent loop detectedExample data

Support Triage called search_kb 38× with the same query.

Run stopped by the loop guard. $4.10 saved

Budget

Workspace, this month

Month to date

$175.35

of $725.00 · forecast $647.46

Top failing agents

Last 7 days

Cost spikeExample data

Research Assistant

3.4×

hourly spend vs 7-day baseline

Trusted by 1,200+ AI teams shipping agents to production

  • Kestrel
  • halcyon
  • Northgate
  • Lumen
  • Quillwork
  • ATLAS
  • Mosaic
  • Tidewell

The problem

A finished run is not a successful run.

Agents fail differently from APIs. They don't throw, they improvise. You need to see what they did, step by step, and stop the risky parts before they happen.

  1. 01

    Silent failures

    The run says “completed”. The agent looped on a tool, gave up and apologised to the customer. Nobody noticed for a week.

    1 in 9 “successful” runs has a hidden error

  2. 02

    Runaway costs

    A retry loop or a 200k-token context turns a $0.05 run into a $4 run. You find out on the invoice.

    Spend spikes are found 11 days late on average

  3. 03

    Unreviewed risky actions

    Refunds, emails to customers, SQL on production. Agents act with your credentials, and nobody signs off.

    63% of teams have no approval step for writes

How it works

From “what happened?” to fixed, in four steps.

Refund Assistant · run traceExample data
Refund Assistant
Plan
lookup_order
refund-policy.md
Decide refund
Human approval: Refunds over $250
issue_refund
send_email
Summarise for agent
AgentLLM callTool callRetrievalHuman approval
01Observe

See every step of every run

LLM calls, tool calls, retrievals and sub-agents on one timeline, with tokens, cost and latency per span.

Loop detected in Support TriageExample data

Likely cause: the help-center MCP server returns an empty page instead of an error.

  • Failing spansearch_kb × 38, identical arguments
  • First seenprompt v14, 2h ago
  • Similar runs3 in the last 24h, all after 07:10
  • Cost of the loop$4.10 across 3 runs

Suggested fix

Add a max-calls guard of 5 for search_kb and treat empty pages as “no results”.

02Diagnose

Find the root cause in minutes

Errors are grouped by cause, with the failing span, similar runs and the prompt version that introduced it.

SDR Outreach

Policy: Bulk email

Medium risk

Send email to 402 contacts

send_email({"sequence":"Q4 expansion: ops leaders", "recipients":402, "from":"[email protected]", "send_at":"now"})

“I found 402 ops leaders matching the ICP and drafted a 3-step sequence. Sending now keeps us inside this week's quota.”

3h 33m left
03Approve

Put a human on risky actions

Policies pause the run before a refund, a bulk email or a SQL write. Approve, reject or edit the arguments.

Research Assistant · OctoberExample data

Month to date

$74.00

of $260.00 · forecast $273.00

$260

Forecast is $13 over. The agent pauses when it hits the limit.

04Control cost

Set budgets that actually stop

Budgets per agent, team or workspace. Alert at 80%, pause at the hard limit, forecast the month.

spans traced last month
48M
spans traced last month
lower LLM spend after 90 days
31%
lower LLM spend after 90 days
median time to root cause
4 min
median time to root cause
ingest uptime, trailing 12 months
99.95%
ingest uptime, trailing 12 months

Traces & replay

Every step, every token, every reason.

  • Nested spans for LLM calls, tools, retrieval and sub-agents
  • Inputs, outputs and the model's reasoning, with PII masked
  • Replay a run from any step with a new prompt or model
Learn more
Research AssistantExample data
Research Assistant
Plan research
Sub-agent: Web researcher
web_search
fetch_url
fetch_url
Extract facts
Sub-agent: Financials researcher
web_search
fetch_url
Summarise filings
Write brief
write_brief

Approval queue

Humans in the loop, without the bottleneck.

  • Policies flag what needs a human: amounts, SQL, MCP servers
  • Approve, reject with a reason, or edit the arguments first
  • SLAs and escalation, so nothing waits all weekend
Learn more
Approvals 4 pendingExample data
  • Issue refund $420 to customer #8812

    Refund Assistant

  • Publish status update for incident #331

    Incident Summarizer

  • Run SQL DELETE on staging.sessions_2024

    Data Cleanup Bot

  • Send email to 402 contacts

    SDR Outreach

Tool-call log

A ledger of everything your agents touched.

  • Every tool invocation with arguments, result and latency
  • Who approved what, and when, for every write
  • Error rates per tool, so flaky integrations show up first
Learn more
Tool callsExample data
TimeToolStatusLatency
09:22:48post_statusHeld—
09:22:36summarize_timelineOK5.4s
09:22:31fetch_alertsOK5.1s
09:16:58query_warehouseOK4.9s
09:10:00lookup_orderOK2.5s
07:14:48search_kbError2.9s

Cost & budgets

Know what each agent costs before the invoice does.

  • Spend by agent, model, team and customer
  • Budgets with soft alerts and hard stops
  • Forecasts, cost-per-run outliers and spike alerts
Learn more

LLM spend, last 30 days

$594.17

Example data
  • Claude Opus 5.5
  • Claude Sonnet 5.5
  • GPT-5
  • Claude Haiku 5.5

And so much more

Everything an agent team needs, nothing it doesn't.

Evals

Turn real runs into test cases and score every prompt version before it ships.

Alerts to Slack

Loops, cost spikes, error rates and SLA breaches, routed to the right channel.

PII redaction

Mask emails, cards and custom patterns before payloads leave your network.

Role-based access

Approvers, developers and viewers. Approval rights per policy, not per workspace.

Prompt versioning

Every prompt change is a version with a diff, an owner and its own metrics.

OpenTelemetry & SDKs

Python and TypeScript SDKs, or send OTel spans from any framework you already use.

Use cases

One control room for every team's agents.

Refunds and replies, with a human on the expensive ones.

Agents

Support Triage Refund Assistant

Example policy

issue_refund where amount > $250 → Support Lead

42%

of tickets resolved with no human touch

Sample result

Integrations

Your whole stack, connected.

OpenAI, Anthropic, LangGraph, Slack, Datadog, and anything else your agents run on.

Explore integrations
  • OpenAI
  • Anthropic
  • Gemini
  • Mistral AI
  • LangChain
  • LangGraph
  • CrewAI
  • MCP
  • OpenTelemetry
  • Slack
  • PagerDuty
  • Linear
  • Datadog
  • Snowflake
  • GitHub
  • Notion

Control cost

Set a budget.
Sleep at night.

Hard limits per agent pause runs before the bill surprises you. Forecasts tell you on day 5 what day 31 will look like.

Customers

“Budgets with a hard stop were the only way finance would let us put agents on the expensive models. Spend is down 30% anyway.”
— Mei Lin Ong, VP Engineering at Quillwork

Security

Built for teams that answer to security reviews.

Visit the trust center

SOC 2 Type II

Audited annually. Report available under NDA from the trust center.

Data residency

Choose US or EU at signup. Enterprise can keep payloads in its own VPC.

SSO & SCIM

SAML and OIDC with any identity provider, with automatic provisioning.

Audit log

Every approval, policy change and API key action, exportable to your SIEM.

Pricing

Start free. Pay per seat when it matters.

Every plan includes traces, approvals and cost tracking. Upgrade for volume, retention and controls.

Starter

For one team getting its first agents into production.

$0free forever
Start free

Includes

  • 100k spans / month
  • 3 agents
  • 7-day trace retention
  • Approval queue (1 policy)
  • Slack alerts
  • Community support

Team

Recommended

For teams running agents that touch money, data or customers.

$39per seat / month
Start 14-day trial

Everything in Starter, plus

  • 2M spans / month included
  • Unlimited agents
  • 30-day retention
  • Unlimited approval policies
  • Budgets with hard limits
  • Evals and prompt versions
  • PagerDuty and webhooks

Enterprise

For organisations with security, scale and compliance needs.

Customcustom contract
Talk to sales

Everything in Team, plus

  • Volume span pricing
  • SSO, SCIM and audit log
  • EU or US data residency
  • Self-hosted collector
  • 1-year retention
  • 99.95% uptime SLA
  • Dedicated support engineer

Compare all features and estimate usage

FAQ

Questions, answered.

Can't find what you're looking for? Write to us and a human answers within a day.

Put your agents in production with a straight face.

Free for your first 100k spans. Set up in ten minutes, see your first trace in two.

No card needed · SOC 2 Type II · Cancel anytime