Labwire is an open protocol that lets AI agents safely control laboratory instruments through a single standardized interface, rather than custom code for each vendor. Every command carries mandatory UCUM units and safety classifications, and runs produce cryptographically signed bundles as tamper-evident proof—essential for auditable lab work. The project is ready to try now with a five-minute quickstart and three realistic simulated instruments, though the protocol spec is still at v0.3 draft and will change before version 1.0.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Labwire, an open-source protocol and SDK, lets AI agents discover and command laboratory instruments through a universal interface. The protocol, now at v0.3 draft, uses JSON-RPC 2.0 over WebSocket, bundles cryptographically signed proof of what each instrument did, and ships with three simulated instruments, an MCP adapter for Claude, and a five-minute Python quickstart.
Why it matters
Today every instrument vendor uses a different dialect, forcing custom integration for each one. Labwire standardizes the conversation so any AI framework can drive any instrument with zero glue code, while mandatory safety interlocks (S0–S3 classes) and UCUM unit codes built into the protocol prevent silent errors and unsafe commands. Signed run bundles give auditable, tamper-evident proof of what happened—critical for regulated lab work.
What to watch
The protocol is still pre-1.0 (v0.3 draft); the wire format will change. Bridges already exist for ophyd (Bluesky) devices and PyLabRobot liquid handlers. Cryptographic operator binding is on the roadmap but not yet shipped. Real hardware drivers are explicitly a non-goal; the project targets the protocol and adapter layer, not reimplementing thousands of vendor drivers.
Labwire is an open protocol and Python SDK designed to let AI agents control laboratory instruments safely and auditably. The core insight is borrowed from MCP (a Claude tool-discovery protocol): instead of custom code for each instrument, an instrument publishes its capabilities as JSON Schema, and any agent framework can drive it with no glue code.
The protocol runs JSON-RPC 2.0 over WebSocket, with an initialize/capability handshake, push-first command lifecycle, and protocol-level safety interlocks. Three design principles stand out. First, every quantity carries a mandatory UCUM unit code (e.g., "uL/min" for flow rate), preventing silent unit errors. Second, every destructive or hazardous action is tagged with a safety class (S0–S3): an S2 action (irreversible, like dispensing) requires operator confirmation; an S3 action (hazardous) requires an operator grant the agent cannot mint. Third, every run produces an ed25519-signed manifest, verified by one CLI command, giving portable tamper-evident proof of what the instrument recorded.
The v0.3 draft adds support for "things"—typed instrument state like a liquid handler's deck—and typed references, so parameters can name a well or site and validate against current state. The protocol is designed so discovery alone leads an agent to the deck with no prompt engineering.
The project ships three realistic simulated instruments (balance, syringe pump, power supply) with latency, noise, drift, and failure modes—not emulations of real hardware, but original models that make the protocol testable in five minutes. The closed-loop demo shows a scripted optimizer tuning heater voltage and reagent flow rate across the three instruments, converging on a hidden yield optimum (87.2% in the example), and then verifying the winning run's signed bundle. `make demo-claude` runs the same loop with Claude as the agent, reading the instrument schemas as MCP tools.
Bridges for ophyd (Bluesky's hardware layer, widely used at synchrotrons) and PyLabRobot (liquid handlers) are already built. The ophyd bridge requires a small YAML annotation to supply units and risk classification, then refuses to serve any device whose quantities lack UCUM codes. The PyLabRobot bridge revealed a design strain: the protocol models actions on quantities well, but ophyd's tree-shaped state has nowhere to live in v0.2. The authors published SPEC-FINDINGS.md listing eight such strains and concrete recommendations for v0.3, rather than smoothing them over.
The protocol v0.3 is still a draft; the wire format will change before 1.0. The project explicitly does not aim to reimplement thousands of vendor drivers, offer fleet control, web UI, auth beyond a stub API key, cloud hosting, or cryptographic operator binding (on the roadmap). It targets the agent-to-instrument protocol layer and the adapter machinery. Real hardware drivers belong to future work or community contribution. The license is Apache-2.0, with patent grant—important for a protocol. Code, spec, examples, and contribution guidelines all live on GitHub under benchwire/labwire.
Labwire addresses a real pain point in autonomous laboratory work: every instrument speaks a different language, forcing scientists and engineers to write bespoke drivers for each one. The protocol models the conversation as a universal schema, borrowing from MCP (Claude's tool-discovery standard) so any agent framework can discover what an instrument can do and command it without glue code. The protocol enforces safety and auditability at the wire level, not as vendor add-ons—mandatory UCUM units prevent silent unit mismatches, S0–S3 safety classes gate irreversible actions behind operator confirmations, and ed25519-signed run bundles create portable, tamper-evident records of what happened.
The project is honest about its scope and limitations. It ships running code, not a specification alone: three realistic simulated instruments, an MCP adapter, SDKs, and a five-minute quickstart prove the design works. Bridges for ophyd (used widely at synchrotron facilities) and PyLabRobot (liquid handlers) are already built. At the same time, the authors are explicit about what is not in v0.2—real hardware drivers, cryptographic operator binding, and fleet control—and have published SPEC-FINDINGS.md with eight concrete design strains found during bridging work, feeding into v0.3. This signals a project more interested in honest iteration than polished claims.
The timing aligns with a broader 2025–2026 shift toward agent-to-instrument protocols. LAP (arXiv:2606.03755) and SCP (arXiv:2512.24189) are converging on similar problems; Labwire borrowed LAP's mandatory unit codes and safety taxonomy with credit. The practical advantage today is that Labwire has running code and SDKs, while LAP remains specification-only.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
