Small payloads matter on a radio link, but the sharper property is that decoding requires no inference — a constrained node recovers the instruction by dictionary lookup, which means no model, no tokenizer and no ambiguity at the receiving end. That makes this interesting for mesh, satellite and embedded contexts and largely irrelevant if every hop in your system is a cloud API call. The validator is where the safety lives: it rejects hallucinated opcodes and missing consequence classes before anything is transmitted.
An encoding standard for agent instructions, with an MCP server in front of it. Instead of JSON over HTTP, instructions are written in a compact grammar that decodes by dictionary lookup rather than inference — so a receiving node recovers the structured instruction without running a model.
- `osmp_compose` turns natural-language instructions into the encoded form, or returns `NL_PASSTHROUGH` when no opcode matches
- An `osmp://system_prompt` resource the agent reads on connect — the README puts it at roughly 390 tokens — teaching it the grammar, dictionary and composition rules
- `osmp_macro_invoke` resolves pre-validated multi-opcode templates, with 16 Meshtastic macros bundled
- `osmp_resolve`, `osmp_batch_resolve` and `osmp_discover` look codes up in the bundled domain corpora — ICD-10-CM, ISO 20022 message definitions, and MITRE ATT&CK
- The MCP server ships 19 tools and wraps the Python SDK
- Encoding is bijective: every instruction compiles to a binary form and decompiles back to the identical text
- A floor guarantee — when the encoded form would be longer than the original, it transmits the original instead, so compression is never negative
`pip install osmp-mcp` for the server, published as osmp-mcp version 1.3.4 over stdio; `claude mcp add osmp -- osmp-mcp` registers it. The SDK is separate and available per language — `pip install osmp`, `npm install osmp-protocol`, `cargo add osmp`, or the Go module. No account and no key. If you use the SDK path rather than MCP, the documented approach is to add the usage doctrine to your model's system prompt and wire encode, decode and validate into your own agent loop.
One command — pip install osmp-mcp
