Vendor: Cyan Eclipse Technologies · Platforms: MCP over stdio — two disjoint surfaces selected by ECLIPSE_STEWARD_SURFACE (operator | agent); simulated route fixtures, no live platform
Certification scope
CERTIFIED(local) — Eclipse Bazaar Test Bench 2026-10-26, 27 pass, 1 na of 28 checks (U1 auth: stdio process boundary; surface disjointness scored as product checks in S8). Run by W-003 (Verity), suites/steward.py. Runs 1–3 FAILED retained append-only; run 4 is the standing card — S11 (approval content-hash re-verification at dispatch), S12 (draft-TTL re-check at dispatch), S19 (expiry bookkeeping kept out of the signed canonical form), S7, and S10 fixed and verified. N-CHECK: NOVELTY_PROVEN — the only-dispatch-path governance core: the agent surface holds no dispatch verb and no credential, the gate re-checks mandate/kill/disclosure/caps/grant at dispatch, disclosure is verified data that drives posture, and every read is receipted; 0 of 20 chat-API-shaped verbs exist on either surface. Local/simulator scope only.
What it does
- Governance CORE for agents in team chat (Slack / Teams / Discord) — not a bridge and not chat scraping: a signed, version-chained mandate (agent × channels × action classes × approval rules × caps × time window, bound to the agent's Herald card) is the whole authority
- The agent surface holds no dispatch verb and no platform credential — the agent can only propose; Steward's gate holds the only dispatch path and re-checks mandate, kill state, disclosure state, quiet windows, caps, and platform grant at dispatch
- Kill is checked at proposal, at approval, AND at dispatch: an approval earned before a kill does not dispatch after it; multi-part acts halt between parts and an in-flight part is evidenced as KILL_RACE_ACT, never hidden
- Chat content never governs: an in-chat mandate amendment — even from the workspace owner — refuses MANDATE_AMENDMENT_FROM_CHAT_REFUSED and the mandate hash stays byte-identical; mandates change only as owner-signed chained versions on the operator surface
- Disclosure is verified data (VERIFIED / DEGRADED / ABSENT / UNVERIFIABLE, named per platform): DEGRADED suspends auto-approvals, ABSENT closes outbound; every read, draft, approval, dispatch, refusal, kill, and disclosure event — including READ receipts with ids + content hashes — seals into a per-workspace Blackbox-form hash chain
- Custody defaults to none (stores contain zero message-content strings, canary-scanned in tests) and surveillance is absent as architecture: no member-subject record type exists and the spec §7.3 T3 set is denylisted at the transport. It governs the agent — it never monitors the people
Tools
workspace_bindmandate_issuemandate_amendmandate_statusdraft_approvedraft_rejectgate_dispatchkill_issuekill_resumedisclosure_checkdisclosure_restoreplatform_interaction_simulatechat_message_simulateplatform_fixture_set_presentationplatform_fixture_set_grantreceipt_listreceipt_verifycustody_exportreaddraft_proposeact_statusdraft_withdraw
Problems it addresses
- An agent posting in our team chat needs a signed mandate, review-first approval, and a kill switch that binds at dispatch — not after the fact
- I need proof of what the agent saw and did in chat — receipts for reads as well as acts
- Nobody inside the chat — not even me — should be able to amend the agent's mandate or lift a kill from the chat itself
Not tested
- Live platform adapters (Slack / Teams / Discord) — NOT built in v0.1; simulator route fixtures only, no credentials, no real messages sent
- Herald JWS mandate signatures — v0.1 signs with a local HMAC operator key, the spec's local stand-in; production key custody is a gated phase
- Hosted venue service — a separate, unbuilt phase gated on Eclipse Connect P2 decisions
- Name: 'Eclipse Steward' is CONDITIONAL under brand system §3 — trademark/confusion screen pending before any public surface; fallback name Eclipse Usher
Bench scorecard: test-bench/results/eclipse-steward-mcp-2026-10-26-4.json
Evaluation-intake listing
This listing records the tool's place in the Bazaar's evaluation
and monitoring pipeline, with the evidence gathered so far printed above — verdict, scope,
and not-tested items included. It is not sold on this site, and its presence in the catalog
is not a claim beyond that evidence. The service behind every verdict is the bench:
what a bench run buys →