The managed MCP gateway · onboarding design partners
Routed, governed and redacted for every AI agent. Ixvane plugs your internal databases, component libraries and browser fleets into AI coding workflows — without handing models the keys.
stdio · sse · streamable-http
MCP made it trivial to give an AI agent a database, a design system or a browser. It also made it trivial to do that with production credentials in a dotfile. Ixvane puts one governed layer between agents and everything they touch.
Developers run local servers with production credentials. Nobody knows how many exist or what they can reach.
~/.cursor/mcp.json "url": "postgres://admin:•••@prod-db"Query results with names, PANs and phone numbers flow straight into third-party model context windows.
tool_result { "pan": "ABCPK4821F", "phone": "+91 98•••" }When an agent drops a table or reads a salary column, security can’t answer who, when, or why.
audit.log — no entries —Agents connect to a single MCP endpoint instead of dozens of servers. Each tool call passes through five stages before it reaches your systems — and again on the way back. Our target: all five in under a millisecond.
one governed MCP endpoint
OIDC or SAML through your IdP. Short-lived tokens per developer, agent and CI job.
Role × tool × argument policies, evaluated in-process — no network hop.
Streaming detectors on requests and responses, with stable per-session tokens.
Pooled upstream connections on in-region edge nodes, with failover.
Every call traced end to end. Export to OpenTelemetry and your SIEM.
These are design previews of the Ixvane console, shared early so you can tell us what you need. Step through the four things a platform team does every day.
Point Ixvane at any MCP server. It discovers the tools, and you decide which ones go live — before any agent can call them.
Every tool call authenticated, authorized, scrubbed and recorded — without slowing your agents down.
target gateway overhead · p99<1 ms
In-region edge nodes terminate MCP sessions, then authorize and route each call in-process. Governance shouldn’t show up in your agent’s latency budget.
India-first detectors — Aadhaar, PAN, UPI, GSTIN — plus emails, cards and secrets, on requests and responses. Values become stable tokens, so agents can still reason over them.
Bring any MCP server, or start from first-party connectors for Postgres, Snowflake, Storybook and Playwright. Tell us which ones you need next — design partners set the connector roadmap.
Request a connector →| Role | query | write | browse | merge |
|---|---|---|---|---|
| admin | allow | allow | allow | approval |
| backend | scoped | approval | allow | allow |
| analyst | scoped | deny | deny | deny |
| vendor | deny | deny | approval | deny |
Okta and Entra groups become roles. Grant per tool, per schema, per row limit — with just-in-time approvals for anything that writes.
Trace each tool call from prompt to upstream and back. See which agents touch which data, what got redacted, and where tokens go.
Ixvane is architected natively around the Anthropic Messages API and the Model Context Protocol. Claude Sonnet and Haiku run beside the gateway — never in its request path — and only ever see redacted payloads.
Haiku grades each prompt for clarity, scope and risk as sessions stream in, so you can see which agents burn tokens or reach for more data than the task needs.
Sonnet condenses long sessions and bulky tool results into compact, faithful summaries, keeping the facts that matter across high-throughput MCP tool executions.
Schema-bound tool calls turn every session into typed records — intents, tools, tokens, redactions and scores — ready for OpenTelemetry, your warehouse or your SIEM.
{ "principal": "ananya.k", "client": "claude-code", "tool": "pg-warehouse.query", "prompt_quality": { "model": "claude-haiku", "score": 0.86, "flags": ["unbounded_date_range"] }, "context": { "model": "claude-sonnet", "turns": 42, "tokens_in": 184210, "tokens_out": 19876, "cache_read_tokens": 15360 }, "pii_redacted": { "email": 2, "pan": 1 }, "gateway_ms": 0.31}Built directly on the Anthropic Messages API, so every model call is versioned, traced and governed like any other tool call.
Native tool calling with strict JSON schemas, so telemetry arrives structured instead of scraped from prose.
Reserved for the hard part: compressing long, tool-heavy sessions without dropping facts.
Tool schemas, policies and redaction profiles are cached as stable prefixes, cutting latency and cost per evaluation.
resource "ixvane_server" "pg" { name = "pg-warehouse" transport = "http" expose = ["query", "describe"] redact = "india-pii"}Point Ixvane at any stdio, SSE or HTTP server — or pick a first-party connector. Credentials stay in your vault.
Roles from your IdP, grants per tool, redaction profiles. Synced from Git and simulated before they ship.
Every IDE and agent gets a single governed MCP URL. Telemetry starts with the first call.
$ ixvane servers add pg-warehouse \ --url $PG_REPLICA_URL --transport http ✓ handshake ok · 9 tools discovered$ ixvane policy apply -f policies/ ✓ 6 roles · 41 grants · redact: india-pii$ ixvane clients token --team backend endpoint https://acme.mcp.ixvane.in/v1 token ixv_live_••••••••••••3f9aYour CISO gets the controls; your developers get the context. Ixvane is being built to process everything in-region and store no raw prompts by default.
YAML policies synced from Git, reviewed in pull requests, and dry-run against recent traffic before they go live.
Immutable, hash-chained audit logs of who called which tool with what scope. Stream them to your SIEM.
In-region processing with customer-managed keys, VPC peering and private connectivity to your databases.
We’re building in the open with a small group of design partners. Timelines are targets, not promises.
We’re looking for 10 engineering teams already running AI coding agents against internal systems. Partners get:
Not yet. We’re in private preview with a small group of design partners. The screens on this page are design previews with sample data — they show what we’re building, not a live system.
A 30-minute call every two weeks and honest feedback. In return you get free access during the preview, a direct line to the founders, and launch pricing locked in.
Tool payloads are designed to be processed in memory on in-region edge nodes. By default only hashed, redacted traces are kept; raw payload capture is off unless you turn it on for a specific server.
Ixvane is built natively on the Anthropic Messages API and the MCP specification. Claude Haiku scores prompt quality in real time, Claude Sonnet with extended thinking compresses long multi-turn sessions, and native tool calling turns each session into structured JSON telemetry, with prompt caching keeping tool schemas and policies warm. These model calls run beside the gateway, not in its request path, and only ever see redacted payloads.
Anything that speaks MCP — Claude Code, Cursor, VS Code agents, Windsurf, Zed, JetBrains and your own agents — over stdio (through a local sidecar), SSE or streamable HTTP.
A self-hosted data plane inside your VPC or on-prem Kubernetes is on the roadmap for general availability. If it’s a requirement for you, tell us — it moves up the list.
We plan to price per developer, with unlimited tool calls and free service accounts for agents and CI. Design partners keep their launch pricing.
We’re onboarding design partners now. Tell us what your agents need to reach.