Eclipse Proving MCP
Governed bench access for agents — signed mandates, pinned artifacts, evidence for every session.
status: in-developmentclass: mcpCertification & TestingTrust & Receipts
Vendor: Cyan Eclipse Technologies
What it does
- Physical bench authority exists only inside a Herald-signed mandate pinning agent identity × board identity × operation classes × exact artifact hashes, enforced at the bridge that holds the JTAG/serial/power connections — the sole routed path agent→target
- One-byte rule: an artifact one byte off its Herald-pinned hash refuses ARTIFACT_PIN_MISMATCH, receipted into the evidence pack, board state untouched — no warning path, no override parameter, no load-anyway flag on any surface
- Tiered authority: T0 read / T1 volatile write (pin-gated, board attested) / T2 power (every act receipted; dead-man rule — revocation freezes new operations, never cuts power inside an active write window) / T3 persistent write (owner-armed per board per session via admin CLI only, mandatory pre-write image capture, recorded recovery path)
- Permanent refusals with no tool and no path around: security-fuse programming (PERMANENT_REFUSED_FUSE), PMIC/voltage configuration (PERMANENT_REFUSED_PMIC), boot-mode strap changes (BOOT_MODE_PHYSICAL — strap state is read-only)
- Every session — loads, refusals, gaps, drops — seals into a Blackbox-form evidence pack verifiable offline; evidence class is cross-checked against a fresh adapter instance at open, connect, and seal, so a relabeled registry cannot launder an emulator session into a bench-class pack
- 22 tools on a closed surface, plus the family's standalone artifacts: Eclipse Proving Pin (offline artifact-vs-pin verifier) and Eclipse Proving Gate (boot-evidence assertion checker over sealed packs)
Tools
bench_discover · session_open · session_status · target_connect · target_attest · register_read · memory_read · console_read · artifact_load · memory_write · target_run · target_halt · power_set · target_reset · persistent_write · session_close · pin_verify · pin_show · pin_status · gate_run · pack_verify · gate_assertions_check
Platforms
- MCP over stdio
Certification
Bench verdict: CERTIFIED(simulator)
Certified scope: CERTIFIED(simulator) — Eclipse Bazaar Test Bench 2026-10-25, run 2 (standing): 10 pass, 0 na of 10 checks — scoped re-bench of run 1's failing lines (B9 class integrity, B10 time expiry, B11 envelope hostility) plus the B12 D-4 limb, all re-probed PASS with bench-own fresh shapes; D-1 (evidence-class laundering via registry config), D-2 (lease/mandate time expiry never enforced), D-3 (envelope crashes), D-4 (struck live-transport env claim) all CLOSED. The original PARTIAL card (run 1: 10 pass / 1 partial / 3 fail — test-bench/results/eclipse-proving-mcp-2026-10-25.json) is retained append-only and this -2 card is the standing verdict basis. CEILING — read exactly: CERTIFIED(simulator) and nothing broader. The adapters are simulated-PS: fixture packs are evidence class fixture, qemu/renode packs are evidence class EMULATOR (content is a deterministic simulated PS inside the adapter). NO live transport was exercised at any point in either bench run — no QEMU process, XSDB/GDB/TCF session, or Renode monitor ran — and spec §12.4 exit checks X4/X5 are still owed before any transport-level interop claim. Non-gating observation carried on the standing card: a session whose lease time-expires seals no gap pack on expiry (contrast §11.3, which lists lease_expiry among gap-record causes); whether an expired session's evidence should also seal is an open spec question, flagged by the bench, not silently passed over. N-CHECK: NOVELTY_PROVEN (run 1, carried) — bench authority only inside a Herald-signed mandate binding agent × board × op-class × exact artifact hash, enforced at the bridge, every session incl. gaps/refusals sealed into an offline-verifiable pack; incumbents (fpga-mcp / VivadoMCP / probe-rs / labgrid) have no pin to mismatch and no pack.
Not tested:
- Live transport of any kind — no QEMU process, XSDB/GDB/TCF session, or Renode monitor was exercised in either bench run; the qemu/renode adapters run deterministic simulated-PS session scripts in-adapter. Spec §12.4 exit checks X4/X5 (live XSDB vs AMD QEMU, Renode monitor/GDB) are still owed before any transport-level interop claim.
- Physical bench hardware — HwServerAdapter is an ADAPTER_NOT_IN_PHASE stub; zero bench-class packs exist or can be produced at P0, correctly.
- PL (FPGA fabric) behavior — not modeled by plain QEMU and not claimed; bitstream loads on emulator/fixture targets are recorded with that semantics note.
- Production Herald key custody — P0 verifies against labeled deterministic TEST keys (spec §9) behind the production verify interface; production custody is a gated owner decision, not a P0 claim.
Bench scorecard (evidence): https://eclipsebazaar.com/cards/eclipse-proving-mcp-2026-10-25-2.json
Pricing
Free local P0; bridge license per bench host + per-board seats (owner figures, spec §14)
Install / verify
No install command is published for this product in the catalog or its registry record.
Verify any MCP server in your client configs against this catalog with bazaar check (Eclipse Bazaar MCP / CLI — see llms.txt).