Velane

Velane is integration lifecycle infrastructure built specifically for AI agents. It gives agents a single MCP connection to discover OAuth-connected accounts, write and test Bun or Python workflow snippets, and promote them through dev, staging, and production environments — all without leaving the chat window. The project is open source under AGPL-3.0, with a hosted cloud option and self-hosting support via Docker Compose or EKS.

What It Is

Velane sits between an AI coding agent (such as Cursor or Claude Code) and the external APIs that agent needs to call. Rather than embedding credentials in code or wiring up OAuth flows manually, the agent calls Velane's MCP tools to discover connected accounts, fetch live API docs, write a snippet, run it in an isolated sandbox, read execution logs, fix errors, and finally publish the workflow to a stable HTTP endpoint. The control plane is written in Go; snippets execute in ephemeral Bun or Python runtimes backed by Redis for queuing and ClickHouse for metrics.

Architecture

Velane's architecture separates a Go-based control plane from an executor pool:

• Control plane — chi router, pgx/v5, JWT and API-key auth, async worker, observability pipeline, and an OAuth proxy backed by Nango (800+ providers)
• Executor pool — isolated Bun (:8081) and Python (:8082) processes; optional VM-boundary isolation via AWS Firecracker (requires KVM)
• Storage layer — Postgres for state, Redis for the queue and warm pool, ClickHouse for per-invocation metrics
• MCP server — exposes list_connections, get_logs, publish_snippet, and related tools so any MCP-compatible agent can manage the full lifecycle

Each invocation runs in an ephemeral sandbox with AES-256-GCM encrypted secrets injected as environment variables and an egress policy (per-tenant IP/CIDR and domain blocklist) enforced inside the executor.

Workflow: From Chat to Production

The three-step agent workflow the homepage describes:

1. Create & Connect — the agent calls list_connections, fetches live API docs for a provider, then writes Bun or Python code saved as a versioned workflow.
2. Run & Debug — the agent invokes the workflow on dev, reads logs via get_logs, and autonomously fixes errors inside the same chat session.
3. Promote — once tests pass, the agent calls publish_snippet with env: "prod", producing a versioned, auditable, rollback-ready HTTP endpoint.

Additional deployment features include canary traffic splitting (route X% of prod traffic to a new version), sync/async/streaming invocation modes, git push-to-deploy (commit SHA tagging), and an embeddable white-label dashboard.

Who It's For

Velane targets two audiences the homepage explicitly names:

• Solo developers who want to call external APIs from an AI agent without maintaining OAuth plumbing, credential management, or backend infrastructure.
• Teams that need shared credential stores, environment promotion, invocation audit trails, and role-based access (invoke / manage / admin scopes) across multiple agents.

A CLI (velane login, snippets push, invoke) is also available for developers who prefer terminal-based workflows alongside the MCP interface.

Update: v0.7.17

The latest GitHub release is v0.7.17, published 2026-07-04, with the repository last pushed on 2026-07-27. The project was created in May 2026 and has been releasing actively since then, signaling early but rapid iteration. The GitHub repository carries an OpenSSF Baseline badge and a CI build-and-push workflow, indicating attention to supply-chain security from the start.