SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI agentsCorvinOS

CorvinOS

CorvinOS is an open-source, self-hosted operating system for AI agents. It claims to deploy agents on every messaging channel, run pipelines, connect organizations via A2A, and comply with the EU AI Act, fully self-hosted. The platform integrates engines, personas, workflows, and runtime-generated tools, enabling full control through browser operation, voice interaction, and messaging bridges such as Telegram, WhatsApp, and Signal, and records all events with a hash-chained audit trail. Its compliance mechanisms are built into the architecture as structural design constraints, each instance has a verifiable identity, and it supports deny-by-default consent for transparency sharing. Targeting a range of use cases from personal assistants to enterprise data center nodes, users can describe outcomes, and the system automatically discovers execution graphs and exports them as portable .awpkg packages.

Open website

About this tool

Purpose

CorvinOS is an open-source AI operating system positioned as a self-hosted platform for AI agents. Its core purposes include: deploying AI agents across various messaging channels, running automation pipelines, connecting different organizations or teams via the A2A protocol, and meeting EU AI Act compliance requirements. It emphasizes "the agent operating system you truly own," with all components self-hosted and compliance as a structural constraint of the architecture rather than an optional mode.

Main Capabilities

Agents and Compute

  • Agentic Compute: The worker owns the loop, the model provides the framework. The agent states the optimization objective and stopping conditions, and the sandboxed worker selects a strategy suited to the problem and executes it, from quick tuning to Bayesian search over an 8 GB dataset, with raw bytes never entering the context window.
  • Multi-engine support: Switch between Claude Code, Codex, OpenCode, local Hermes/Ollama, and GitHub Copilot without changing the bridge. Built-in adaptive Haiku/Sonnet model selection.
  • Recursive delegation: The manager engine plans tasks, delegates parts to run on parallel workers, scores each result, and iterates toward a loss objective. When a capability is missing, workers can forge the required tools or skills at runtime (Forge and SkillForge).

Tool and Skill Generation

  • Forge: After the worker calls forge.create_tool() , the generated tool is immediately available to all subsequent workers; tools are schema-bound, sandboxed, invoked via MCP, and have a four-scope workspace hierarchy.
  • SkillForge: After the worker calls skill.create() , the skill is injected in the next round. Skills exist as markdown, undergo automatic scoring, promotion gating, and an injection linter.

Compliance and Security

  • Each instance carries a verifiable identity: a locally generated Ed25519 key becomes a Corvin Labs-signed certificate upon license activation, verifiable offline, optionally bound to machine hardware.
  • A one-time AI nature disclosure per user, structurally locked and unable to be disabled via configuration.
  • Deny-by-default transcript sharing consent, per user, with a TTL cap, revalidated at consumption time.
  • Hash-chained audit: every event is appended to a SHA-256 linked chain, tampering invalidates it, and it can be verified offline.
  • Compliance-zone routing and egress locking structurally prevent data from leaving the licensed region.

Data and Queries

  • Data firewall: Once a dataset is registered, it is presented in two ways—a redacted snapshot for the model, and real bytes for the sandbox. The model sees a 22-character opaque data_handle, schema and redacted samples (capped at 4,000 tokens), statistics computed via HyperLogLog and P5/P95; the sandbox receives the real file path, read-only bound to bwrap, and can read line by line.
  • DSI protocol: Points to an existing database, the model queries by name, and raw rows do not move.
  • Conversation memory: FTS5 SQLite recall, PII-redacted index, via /forget Implement GDPR Article 17 erasure.

Workflow and Packaging

  • AWP Workflow: Describe outcomes in natural language, the manager delegates execution, and the execution graph discovers itself. Graphs can be saved as repeatable workflows and exported as a single portable .awpkg package.
  • .awpkg: A signed ZIP containing logic, tools, skills, personas, and data bindings. On installation, it is decompressed after eight fail-closed checks, and tools, skills, personas, and delegation logic take effect together.

Communication and Control

  • Message Bridges: Native daemons for Discord, Telegram, WhatsApp, Slack, Email, Teams, and Signal. Supports hot-reload settings, per-chat configuration, and rate limiting. Each bridge is a complete control plane, capable of running the entire OS from a Telegram, WhatsApp, or Signal thread.
  • Voice to Action: Voice requests are transcribed locally (by default), routed, executed, and answered back in a listening-friendly way. Audio is deleted immediately after return, and only metadata enters the audit chain.
  • Browser Control: Drives real Chromium, can navigate, fill in, click, and read, while allowing real-time viewing and approval of sensitive operations. Acts by element index rather than pixels; keys come from the vault and do not enter the model or audit logs.

Extensions and Integrations

  • Install external MCP servers from the catalog, add provider layers via the Extension API, or connect custom compute engines.
  • Seven extension surfaces, most hot-reload on the next message, and every event is written to a hash-chain audit log.
  • corvin.* The core remains cryptographically locked—extensions can only add guardrails, not weaken them.
  • Community-built content lives in an open MIT-licensed marketplace.

How to Use

CorvinOS is a self-hosted system installed on the user's own platform. It can be interacted with via message bridges, voice, or a browser interface. Users can describe tasks in natural language, and the manager delegates them to workers for execution. Workflows can be saved as repeatable graphs and exported as .awpkg packages, installable to any tenant with a single command. Specific installation commands and platform support information are not yet provided in the supplied materials.

Use Cases

The materials list nine ways to use it, including: personal chat assistant (Telegram, WhatsApp, Signal, or web, text or voice); retrieval-based Q&A over one's own notes and files; voice messages and voice replies; querying data warehouses such as Postgres or Snowflake via natural language; automating repetitive work (multi-step AWP pipelines, run on demand or on a schedule); customer support (based on one's own documents, with redaction and auditing); delegation between teams via HMAC-signed envelopes (legal, support, data, and other teams each run agents, with peer scope as Observer or Executor); and organizational data center nodes running 100-iteration scans, tuning, and backtesting.

Limitations

The limitations explicitly stated in the materials include: extensions cannot weaken the corvin.* core's security properties;.awpkg eight fail-closed checks before decompression; key values in data bindings remain in the vault and do not propagate with the package; audit logs contain only metadata and do not include transcription text, input keys, or raw rows; voice audio is transcribed locally by default and deleted immediately; each user's one-time AI nature disclosure cannot be disabled; and transcription sharing consent defaults to deny with a TTL cap. Other potential limitations (such as platform requirements, resource consumption, the full list of supported models, etc.) are not yet provided in the given materials.

Related tools

Browse category