The AI agent ecosystem has long been obsessed with orchestration inside a single process, but the new open-source protocol Mitra shifts the focus to the harder problem: how two agents that have never met, don't share a runtime, and don't trust each other can safely exchange work. Released on October 20, 2026, on GitHub by user mrpkk, Mitra is a small, MIT-licensed Python package with no required dependencies. It introduces a standard for agent-to-agent communication that includes budgets, mandates, and cryptographic signatures, aiming to replace fragile chat-like interactions with verifiable, bounded transactions.

The Envelope Model: Budgets Travel With the Message

At the core of Mitra is the concept that a message between agents is an envelope, not a chat line. This envelope carries critical metadata such as the goal, task, recipient, token budget, price in USDC, and a mandate with a maximum cost in USD. By embedding these constraints directly into the message structure, the protocol ensures that context assembly is a hard constraint rather than a hopeful suggestion. If an agent receives a message, it knows the exact limits before it begins processing, preventing runaway costs and ensuring that refusals are explicit. For example, if a mandate is exceeded, the system returns a specific error like 'amount: 50.0 above the mandate of 0.5' instead of a generic boolean, allowing the receiving agent to act on the reason for refusal.

Three-Tier Memory and Offline Verification

Mitra implements a three-tier memory architecture to prevent context degradation, a common failure mode where agents become confidently wrong due to silent truncation. The tiers include 'hot' memory in the envelope queue, 'warm' memory in a SQLite database with full-text search, and 'cold' memory for artifacts addressed by SHA256 hash. The context window never grows; instead, the store does, with items fetched only if they fit the budget. Furthermore, all envelopes are signed using Ed25519, with the public key traveling inside the envelope itself. This allows any skeptic to verify the signature offline without trusting a central hub or holding a secret, a crucial feature for decentralized agent networks.

Conformance Suite Ensures Interoperability

Understanding that protocols only succeed through widespread adoption, the Mitra repository ships with a rigorous conformance suite designed to test any implementation. The suite includes ten specific checks, such as verifying that signatures are valid, tampering invalidates them, token budgets are respected, and expired envelopes are not delivered. Developers can run the suite against their own hub implementations using a simple pytest command, ensuring compatibility with the reference implementation. This focus on testable compliance aims to prevent fragmentation, allowing agents built on different hubs to communicate seamlessly if they pass the same conformance tests.

Key Takeaways

  • Mitra is a lightweight, MIT-licensed Python protocol designed for agent-to-agent communication, not intra-process orchestration.
  • The protocol uses signed envelopes with embedded budgets and mandates to enforce hard constraints on cost and context.
  • A three-tier memory system (hot, warm, cold) prevents silent context truncation, a major source of agent hallucinations.
  • The repository includes a 10-check conformance suite to ensure interoperability between different agent implementations.
  • Mitra is explicitly not a runtime or hosted platform, but a carrier protocol that can integrate with existing standards like MCP and A2A.

The Bottom Line

For builders tired of debugging agent loops that silently drop context or blow out budgets, Mitra’s envelope model offers a pragmatic, verifiable alternative to hand-rolled JSON exchanges. It’s small enough to audit, strict enough to be useful, and open enough to actually get adopted.