Netra

Netra sits around an agent rather than inside it. The SDK auto-instruments the libraries an app already calls, and every prompt, model call, tool invocation and retrieval arrives as an OpenTelemetry span that can be replayed. The vendor's framing of the problem is the failure that logs miss: a run that returns 200 and still answers wrongly, described on the home page as "including the ones that return 200 and still get it wrong".

On top of that trace store sit six more surfaces, each with its own docs section: prompt management with versioned prompts, offline evaluation against datasets, multi-turn simulation with personas and goals, red teaming, online evaluation of live traffic, and clustering of production sessions by intent. The docs index describes evaluation as "datasets + evaluators (LLM-as-judge and code) + test runs for regression tracking", which is the shape a team already running CI will recognise.

Voice is a first-class target, not a late addition. The integration list covers LiveKit and Vapi as voice frameworks and Deepgram, ElevenLabs and Cartesia for speech, described as "Evaluate and simulate voice agents through automated call testing and telemetry ingestion".

Three tiers, two of them priced. Free is "$0 forever" with 5,000 spans a month, 25 test runs, 100 evaluator executions, 5 users and 7-day retention. Pro is $39 a month for 200,000 spans, 100 test runs, 5,000 evaluator executions, 50 users and 30-day retention. Custom is "let's talk".

The number to model is spans, because a single agent request produces many of them. Pro publishes its overage openly at "+$5 per 100k after", which is more than most observability vendors disclose and makes a rough bill calculable before signup. Seats are not metered separately at either paid level.

No monthly and annual billing control exists on the pricing page: the plan cards state one term each, and the page carries no annual toggle, pill or discount sentence. Several capabilities are tier-gated rather than priced, and one of them is a headline feature: red teaming, Agent Insights, unlimited test runs and custom data retention are listed only in the Custom column.

The top tier is where the deployment options live too. On-premise and multi-region deployments, an isolated database, enterprise SSO and dedicated Slack support are all Custom-only, so a team that needs its data in its own environment is in a sales conversation rather than a self-serve one.

Netra runs a remote MCP server at region-specific endpoints, authenticated with an x-api-key header, and the docs enumerate "21 tools" grouped into observability, datasets, evaluators and test runs. Setup snippets are published for Cursor, Claude Code, Windsurf, Antigravity and Codex.

What the tools do is worth reading closely, because the docs are unusually candid about the boundary. MCP covers evaluation setup and result inspection; the agent under test still has to be executed separately. The docs say creating a test run "registers the run but does not execute your agent or trigger evaluators on its own". An assistant can therefore query traces, build a dataset and read scores from the IDE, but the run loop belongs to the SDK.

Alongside it the vendor publishes Netra skills for coding agents, offered on the home page as a single paste-in prompt that works with "Claude Code, Cursor, Codex, and any skills-ready agent".

Two things carry an Apache 2.0 licence. The Python and TypeScript SDKs are public repositories under KeyValueSoftwareSystems, and so is Agent OPFOR, the red teaming engine, which the home page links with a star count and labels "Open source red teaming". Its README calls it "Open-source adversary emulation for AI agents, LLM apps, and MCP servers" and maps its suites to OWASP LLM, Agentic AI, MCP and API Security top tens plus an EU AI Act bias suite.

OPFOR is usable without Netra at all: a CLI, a browser extension, an SDK, IDE skills, and a mode where it runs as an MCP server so a coding assistant can red team other agents. It also integrates with Langfuse, a competing trace backend, which is a fair signal that it is a real project rather than a lead magnet.

The platform itself is not published. Netra is sold as hosted cloud, dedicated cloud or on-premise, and the on-premise route is a Custom-tier commercial arrangement rather than source anyone can read.

Netra comes out of KeyValue, an Indian product engineering company. The about page says the team has been "building and shipping software for 10 years" and deployed more than 25 agents in production for clients before building the platform, which matches the OPFOR README's claim of 130 products for 90 startups.

The security claims are named rather than gestured at. The FAQ states Netra is "SOC 2 Type II certified and HIPAA and GDPR compliant" and adds tenant isolation, encryption in transit and at rest, PII detection and masking, and US and EU residency with no cross-region sharing. Residency is visible in the product rather than only in prose: separate US and EU app hosts, OTLP endpoints and MCP endpoints are documented, and the signup link on the site points at the EU host. No SOC 2 report, trust centre page or standalone DPA is published for a buyer to read, so the certification rests on the FAQ sentence.

The reliability and efficiency figures on the home page (1B+ spans a month, 99.99% uptime, 95% reduction in mean time to resolution, 75% reduction in AI spend) are vendor-stated, with named customer quotes from Pencil, Aartha AI and Skydda attached but no independent audit behind the percentages.

It fits a team that already has agents in production, writes the code, and wants traces, eval gates and simulation in one place instead of three tools. The migration pages for Helicone and Portkey suggest the vendor expects buyers to arrive from a narrower tool, and the 14 model providers, 10 orchestration frameworks and 8 vector databases in the integration list cover most current Python and TypeScript stacks.

It does not fit a buyer who wants reliability without engineering. Nothing happens until an SDK is installed and initialized, and the evaluation loop runs from code. A product or operations owner cannot self-serve value here.

Two other mismatches are worth naming. A team whose main requirement is adversarial testing can take Agent OPFOR for free and skip the platform, since red teaming inside Netra is Custom-tier. And a small team on Free will find 5,000 spans and 7-day retention enough to prototype and not enough to debug a live incident from last week, which is the use case the product is sold on.