Product/Overview

One neutral control plane in front of every AI call

Apps and agents hold one obstruo key. Every request passes through the control plane, redacted, routed, enforced and logged, before it reaches any provider or MCP server.

The control plane

One neutral layer in front of every call

Apps and agents hold one obstruo key. Every request passes through the control plane, redacted, routed, enforced, and logged, before it reaches any provider or MCP server.

Your apps and agents
Support copilot
Treasury agent
Internal tools
Control plane
Redact
Route
Enforce
Audit
Providers and MCP servers
OpenAI, Anthropic & othermodels
Google, Microsoft & othermodels
MCP serverstools
In-path enforcement

Active control, not only observability

Because every call already passes through the control plane, obstruo can act on it in path. No queue, no agent to wake, no redeploy. Three kinds of control, one plane.

LIVE
Negative control
redact, block

Strip or stop what must not pass: PII, secrets, jailbreaks, prompt injection, before it reaches the model.

BUILDING
Positive control
inject, mandate

Require what policy says must be present: context, disclaimers, allowed tools, on the way through.

COMING SOON
Deferred control
hold, escalate

When policy says a human must decide, hold the action, escalate with context, and emit a signed Article 14 record.

The product

One console for the whole AI surface

From keys and the MCP proxy to the governance registries your agents will need. What the control plane enforces on each call, redaction, guardrails, limits and budgets, now lives on runtime controls. Some surfaces are live today, others are in preview, with the governance model already wired into the control plane.

ObservabilityLIVE

Every call, with the evidence to prove it

One audit trail across every provider and agent: the full call chain, the routing decision, and the redaction state preserved exactly as it was at request time. Query it, export it, and retain it for years.

  • Auditor view and CSV export
  • Retention up to 2,555 days
  • Filter by key, model, status, or PII state
Audit logExport CSV
Any key Any model Any status Has redactions
TIMECLIENTMODELPIISTATUS
17:05:2910.60.0.16obstruo-chat3 redacted200
17:04:5810.60.0.04obstruo-chatnonefailover
17:02:1110.60.0.09obstruo-embedding1 redacted200
MCP endpoint, eu-agents-proxyGOVERNED
3 MCP servers 1 endpoint 7 of 12 tools allowed
TOOLSERVERACCESS
github.create_issuegithub-mcpAllow
github.delete_repo github-mcpDeny
jira.search_issuesjira-mcpAllow
postgres.querypostgres-mcpAllow
postgres.execute postgres-mcpDeny
MCP proxyLIVE

Put every MCP server behind a governed proxy

Register your customer MCP servers as upstreams, then expose them to agents through obstruo MCP endpoints. It is many-to-many: one server can back many endpoints, one endpoint can compose many servers, and you allow or deny each tool per proxy.

  • Per-tool allow and deny on every endpoint
  • Destructive write and delete tools flagged before an agent connects
  • Trust, transport, and auth posture on every server
Integration

One key. One endpoint.

Point your existing SDK at obstruo and change nothing else. Provider keys live in the vault, redaction and routing apply in path, and the audit record is written before a single token leaves your network.

Drop-in OpenAI-compatible API
Logical models resolve to the real provider
Around 20ms added latency, p50
from openai import OpenAI

client = OpenAI(
    base_url="https://api.obstruo.ai/northwind/v1",
    api_key=OBSTRUO_KEY,   # one key, provider keys live in obstruo
)

resp = client.chat.completions.create(
    model="obstruo-chat",   # logical model, obstruo picks the provider
    messages=[{"role": "user", "content": prompt}],
)
# PII redacted, call routed, decision written to the audit log
Why obstruo

Built for control, not lock-in

Provider independence

Route across models, providers, and regions by policy and cost. Failover when one degrades. No governance locked into a single vendor ecosystem.

EU data residency

EU-resident by default, in EU regions only. Pin traffic to in-jurisdiction endpoints and keep processing where your regulators expect it.

Active governance

Policy is enforced in path on every call, not reported on after the fact. Redact, route, hold, or block, and emit signed evidence as it happens.

One registry for the AI surface

Agents, MCP servers, and skills in one place: owners, risk, data sources, and per-tool permissions, discovered from real live traffic.

Your AI. Your data. Your control.

Start free on one project, then make the same policy executable across the organisation.

See obstruo in action See pricing