Skip to main content
Glama
CYAN-ECLIPSE-TECHNOLOGIES-LLC

Cyan Eclipse Tether MCP

Official

Eclipse Tether MCP

Powered by Cyan Eclipse Technologies — https://cetech.ai

Eclipse Tether — governed robot task authority — signed mandates, bounded envelopes, evidence for every task.

Agents can already command robots through public MCP wrappers whose authority model is: whoever holds the robot password holds the robot. Tether is the mandate layer between an agent and a moving machine: physical task authority is exercised only inside a Herald-signed mandate pinning which agent may dispatch which pinned, human-authored task templates to which robot serials, inside which workspace envelope, under which speed/force ceilings and occupancy class, until when — enforced at a site bridge that is the only routed path to the robot — and every session, including its refusals, gaps, envelope exits, and human overrides, is sealed into a Blackbox-form evidence pack a third party can verify offline.

Spec: ../eclipse-tether-spec-20261025.md.

Status — honest labels

v0.1.0, local stdio transport, Python stdlib only. P0 = simulator only: ROS 2 + Gazebo + UR-class arm represented by an in-process sim robot behind the bridge (planner: sim, class: fixture). No real motion, no hardware calls, no vendor SDK, no network surface. Liability counsel + insurance + vendor-terms reviews gate P1 real motion only (spec §10.4) — they do not gate this build. Certification ceiling for P0 is CERTIFIED(simulator), granted only by a Verity bench run, never by the builder.

Platform positions (spec §4, binding): Tesla Optimus = published wait position — closed platform, no SDK/API exists; never a logo wall. Spot position = governance + evidence for SDK-direct and multi-vendor sites — not Orbit dispatch (Orbit 5.2 is Boston Dynamics' own vendor-cloud MCP). Standards landscape: ISO 10218-1/-2:2025, ANSI/A3 R15.06-2025, ANSI/RIA R15.08 as spec'd.

Related MCP server: MELRA

Safety law (spec §3, binding on every surface)

  1. A mandate never substitutes for a site risk assessment.

  2. Ceilings are outputs of a human risk assessment, entered by humans; no agent proposes/computes/raises a ceiling. Shipped sim defaults are refusal-tight placeholders, labeled as such.

  3. No T2 direct motion in v0/v1 — no velocity/joint/torque/trajectory surface exists at any tier, under any mandate.

  4. Occupancy is human-declared (UNOCCUPIED / SUPERVISED / SHARED); unknown = T0 observe-only.

  5. Human stop authority is absolute: e-stop / lease grab / local control ends the session (HUMAN_OVERRIDE, terminal); agent resume refuses until human clearance + re-attestation.

  6. Vendor safety systems are outside the type system — absent, not permission-gated. Mandate ceilings may only tighten vendor-rated limits; lower value wins, disagreement receipted.

Surfaces — 22 tools (spec §6)

Selected by ECLIPSE_TETHER_SURFACE; disjoint dispatch tables; a verb absent from a surface returns the generic unknown-tool refusal, byte-identical in shape to gibberish (no existence oracle).

Surface

Tools

agent (9)

fleet_discover session_open task_templates_list task_dispatch task_status task_abort telemetry_read image_read session_close

operator (6)

mandate_register mandate_revoke template_register template_revoke site_freeze_set session_reattest

tools (7, offline)

pack_verify pack_render pack_compare gauge_check_mandate gauge_check_template gauge_check_pair gauge_limits_list

Operator destructive verbs use the house verified-approval pattern: the operator approval signature is re-verified against the operator key at use time. Simulated vendor endpoints (e-stop, lease grab, heartbeat, evidence writer) are driven by fixtures/sim_control.json — the P0 fixture form of §10.2 step 6. Bench-only fixture clock: tools accept at_ts (ISO-8601 with offset), labeled class: fixture wherever it surfaces.

Refusal anchors

ENVELOPE_VIOLATION is a pre-motion refusal (robot never moves, receipted, sealed in the pack). MANDATE_EXPIRED_STOP seals a mid-task expiry with final pose. HUMAN_OVERRIDE is terminal for the session and is also the refusal class for every agent resume attempt on the same mandate until human clearance. Full closed vocabulary: spec §6.4, mirrored in tether_engine.REFUSALS.

Run

python3 fixtures/make_fixtures.py   # mint TEST fixture keys + sim site
python3 smoke.py                    # 22-tool surface enumeration
python3 tests/test_tether.py        # builder regression tests (not certification)
ECLIPSE_TETHER_SURFACE=agent python3 server.py

Not tested (manifest, verbatim from spec §13)

real hardware of any class · real vendor SDKs (Spot/Unitree/UR-RTDE) · real e-stop/lease hardware · Isaac Sim · multi-robot sessions · hosted/multi-site · T2 (refused by design, §3) · closed platforms: Tesla Optimus, Figure, Apptronik, Atlas, NEO (no surface exists, §4.1) · Orbit dispatch (vendor's, not ours, §4.2).


Eclipse Tether MCP · © 2026 Cyan Eclipse Technologies LLC · cetech.ai

Related MCP Connectors

Related MCP Servers

  • A
    license
    B
    quality
    A
    maintenance
    Open-source local MCP runtime. MELRA checks tool calls against policy, pauses scoped mutations for approval, verifies outcomes, and reconciles uncertain writes. npm: @melra/cli. Alpha; Node 22.13+ or 23.4+.
    11
    13 npm
    13
    AGPL 3.0
  • A
    license
    Not graded
    quality
    B
    maintenance
    Provides a governed execution boundary for AI agents, enforcing deterministic policy, one-shot human approvals, and signed receipts for MCP tool effects.
    MIT
  • A
    license
    B
    quality
    A
    maintenance
    An MCP server that lets an AI agent operate an owner-connected robot only through the owner's pre-approved, bounded actions, enforcing rate and motion limits, leases, replay and expiry checks, owner-only confirmation and stop controls, and a tamper-evident audit log. It also grades any robot connector against a hijacked agent via a built-in safety benchmark.
    6
    Apache 2.0