MeterCall (also called RailCall in the official documentation) is a local-first AI workflow and agent execution platform positioned as a "governed AI execution layer": AI handles thinking and drafting, while RailCall handles secure execution on the local machine and leaves a verifiable signed receipt at every step. Its core proposition is to let teams use AI with trust to handle sensitive business such as finance, CRM, and email.
In terms of use, RailCall is aimed at scenarios where AI agents or workflows need to touch real business data but must also meet security and compliance requirements, such as customer CRM automation, invoicing and collections, settlement, and other processes involving funds or customer records, as well as scenarios where compliance departments require that "agents must not touch the production environment."
Main capabilities include: First, local execution. All write operations are completed in local memory, and data, keys, and code do not leave the local machine. Second, dry-run by default and human approval. Each write is first staged with zero side effects, generating a readable signed diff, and is only actually executed after human approval. Third, verifiable receipts. Each run generates an Ed25519-signed receipt locally, which can be independently verified offline; the signature of a tampered receipt will fail. Fourth, a local key vault. Keys are stored locally in a vault with 0600 permissions and do not pass through the cloud. Fifth, isolation verification. During processing, real-time lsof socket scanning ensures zero data egress; the only outbound connection is an instantaneous billing pulse carrying the client SHA-256 key hash and a random nonce. Sixth, MCP server support. It can serve as a local Model Context Protocol server and work with AI clients such as Cursor, VS Code (Cline/Roo), and Claude Desktop/Code, with the AI client responsible for "thinking" and RailCall responsible for "execution." Seventh, a standalone terminal CLI. It supports login, audit, and verification via the command line, for example railcall login <key>、railcall audit <file.csv>、railcall verify <receipt.json>; Eighth, a pre-audited component library. Programs are composed of components with clean licenses and pinned versions, rather than freely generated code. Ninth, approval policies. The boundary between automatic and human approval can be set, with the strictest setting by default, where all writes wait for human confirmation; policy changes themselves also require approval and are logged. Tenth, dual-mode integration. n8n, Zapier, and custom pipelines can be retained, routing sensitive data into RailCall for local processing and then returning structured results.
In terms of usage, users can install the local binary via a one-line install command curl -fsSL https://railcall.ai/install.sh | bash and use it after logging in. For development, it provides the @airlock.flow decorator, which can wrap any Python or TypeScript function into an audited, metered, and receipted process with just one line of code. Billing is per process rather than per seat, at $0.01 per governed process, including building, auditing, or explaining an operation once; composition, dry-runs, and rebuilding after drift are not charged again. The free tier is 100 processes, with no credit card required. Teams share one balance, and inviting members is free.
Applicable scenarios cover regulated work such as finance, healthcare, and enterprise CRM, as well as teams that need client account automation, secure fund writes, and compliance approval. In terms of limitations, the official documentation clearly states that heavy computation runs on local hardware and requires a local environment; the current strictest setting is that all writes wait for human approval, and approval policies will be introduced in later versions; the enterprise-grade air-gapped layer is directional planning, with no specific date promised; payments currently support credit cards via Stripe, with USDC and autonomous budget wallets coming soon; enterprise contracts can provide SSO/SCIM, scoped keys, and a DPA. The product is led by Patrick Linden and others, and it explicitly does not accept outside venture capital.
