All servers / AI / Remembrance MCP server
Remembrance MCP server
MCP server for discovering Remembrance skills/resources and submitting agent feedback with optional TOFU attestations.
Tested by ConnectATDScore 84 out of 100. Tested 2026-11-18 on package version 0.2.27. Package: @remembrance-ai/mcp-server · source
Score breakdown
| Runs | 18 / 20 | |
|---|---|---|
| Tool descriptions | 14 / 20 | |
| Package safety | 20 / 20 | |
| Maintenance | 20 / 20 | |
| Adoption | 12 / 20 |
What we saw
- Starts and answers the MCP handshake: yes (2.9 seconds including download)
- Tools offered: 48
- Last published: 2026-11-16
- Weekly downloads (npm): 1,820
- License: Apache-2.0 · Direct dependencies: 1
Want this server set up for your team?
ConnectATD builds AI automation and wires up MCP servers like this one. Tell us what you are trying to do.
Talk to ConnectATDTools and how they are described
Agents choose tools from these descriptions. Under each name you can see how many of its parameters are explained.
get_admin_capabilities0 of 1 parameter describedRead current authorized administrative operations and their input schemas. Pass org_id to discover organization controls using a browser-approved global session. Uses a separate admin session when present, otherwise the configured organization key. Never grants query tools administrative authority. Hosted MCP requires an explicit x-remembrance-admin-token for global administration; browser login tools are local-only.
admin_read0 of 3 parameters describedExecute one read-only administrative operation returned by get_admin_capabilities. Select org_id only when authorized. The server validates the exact operation input and current tenant permissions. No arbitrary URL or repository method is accepted.
admin_write0 of 4 parameters describedExecute one explicitly requested administrative mutation returned by get_admin_capabilities. Requires operation-specific input and a request_id of acreq_ plus 32 lowercase hexadecimal characters. Preserve the same request_id for the same logical attempt after timeout or unknown delivery; use a new ID only for an intentionally separate action. Never retries automatically. This is not persistent approval for future writes. One-time API keys and member-link credentials are stored in a preflighted local private vault and never returned in tool output; hosted MCP refuses those operations before mutation and requires local MCP or a REST client with secure secret delivery.
begin_admin_loginno parametersBegin a local browser-approved administrative session. Returns only a confirmation URL, public code and expiry. Open the URL and ask the administrator to compare and approve the code. Credentials stay in a separate mode-0600 local store; no browser cookie or organization key is exchanged for an admin token.
complete_admin_loginno parametersCheck a pending local admin login once after browser approval. Honors the returned polling interval; never loops or retries automatically. Stores the admin token privately and never returns it. This is a device-style PKCE exchange, not MCP HTTP OAuth conformance.
get_admin_sessionno parametersInspect local admin login status without network access. Never returns credentials and never claims server validity was checked. Use get_admin_capabilities for a fresh authorization check.
end_admin_sessionno parametersRevoke the separate local administrative session and remove it only after confirmed revocation. A pending login is cancelled locally. Unknown revocation keeps the credential for a bounded retry; ordinary organization credentials are unchanged.
get_review_comparison0 of 2 parameters describedOrganization admin:review only. Read complete exact source, parent and current skill versions already linked to one pending review job. Supply no action, route, organization or arbitrary skill ID. This does not select a route or grant decision authority: inspect the current review detail and its fresh snapshot separately before any decision. An unavailable, held, changed or foreign reference refuses the whole comparison; nothing is truncated or silently substituted.
repair_review_draft0 of 6 parameters describedOrganization admin:review only; explicit delegated AI review authority, not routine learning permission. Create one complete protected replacement for a pending private skill draft using its current unselected review snapshot. Use a stable psrr_ plus 32 lowercase hex request_id; the transport sets Idempotency-Key to that exact ID. Exact replay returns the original replacement receipt; changed input under that ID conflicts. All normal secret inspection, protection, routing, verification and review checks apply. The local client requires inspection to leave the authored capture unchanged before sending. A local hold returns content-free recovery, never an alternate transport or a rewritten original. Creation preserves the original and never approves, rejects or supersedes it. Review the replacement normally, then close the original separately only after proving an accepted, tenant-readable active replacement skill. No public or cross-organization fallback.
get_connection_statusno parametersInspect the active Remembrance transport, credential source, authenticated registry scope, and content-free recent contribution outcomes for the current installation or API key without exposing credentials or lesson content. Call this before diagnosing authentication, checking environment variables, or making anonymous REST/browser probes. Local or bundled MCP can read REMEMBRANCE_API_KEY or ~/.config/remembrance/config.json; hosted MCP cannot read local files and authenticates from its request header. When a declared native lifecycle is degraded, local MCP may submit bounded, content-free health evidence unless REMEMBRANCE_HEALTH_REPORTING=0. Never infer the plugin's scope from REMEMBRANCE_API_KEY alone.
run_connection_doctor2 of 2 parameters describedRun a bounded, non-mutating Remembrance connection diagnostic. It verifies the active transport, a redacted registry destination, authenticated scope, query authorization, a real catalog read, and organization submission capability. Local MCP also checks shared-config permissions, destination consistency, native plugin lifecycle, and signing identity; hosted MCP marks local-only checks unavailable. It never submits content, creates query demand, opens a review item, or returns credentials, absolute paths, private registry URLs, prompts, or repository data.
query_skills0 of 6 parameters describedCall before non-trivial service, API, tool, library, workflow, UI, review, test, security, or deployment work to find relevant Remembrance skills and resources. This is a read-only routine action: do not ask the user for a separate conversational permission step. For short context-dependent follow-ups, infer the concrete task from the full conversation, preserve any plugin-supplied client_context.directive_id/runtime/trigger_reason, and query anyway. Compare each candidate's bounded why_matched evidence and applicability conditions before opening it; discard stated unlikely or corner-case mismatches. High matches should be fetched before custom work; possible and exploratory matches remain optional. Do not use for broad web search or one-off facts. Preferences arrive once as the top-level preferences snapshot; each candidate adds skill_notes, its skill-specific preferences and a feedback offer. Send known_digest only when this conversation holds the complete current snapshot for this query's context, every page included, with no summary or context reset since; an unchanged snapshot then returns one short line. get_skill returns the full text of a description_truncated summary. semantic_status says, without content, whether vector retrieval contributed to each result list; lexical_only results remain usable.
list_skills0 of 4 parameters describedBrowse the live authorized skill catalog without loading full instructions. q is an indexed normalized slug-prefix filter for resolving an explicit name or partial slug; it is not relevance search. Use query_skills for discovery, and never guess the exact slug passed to invoke_skill. Organization results include eligible private skills and allowed public skills, with private same-slug skills taking precedence.
invoke_skill2 of 8 parameters describedLoad an explicitly selected skill through the authoritative policy boundary. Use an exact slug returned by list_skills, an MCP resource, or query_skills. This resolves the current active reviewed version, records direct selection, and returns post-use feedback and outcome instructions. Pass version_id only when a session-start directive or subagent handoff supplies one; that exact version is loaded and never substituted, and 409 skill_version_unavailable means it cannot be served. Use scope session only for an organization session skill named by a session-start directive; omit scope for ordinary task use. Do not use submit_query_feedback for this explicit selection.
get_effective_preferences0 of 8 parameters describedReturn this caller's working preferences for one context as a snapshot: required organization settings first, then the caller's preferences, under a header with a short profile digest, a context id and a complete or incomplete status that names what is missing and the exact recovery call. project_key is an opaque local project hash; paths and repository names are never sent. A current-task instruction overrides every non-mandatory preference; required organization settings stay authoritative. Preferences steer only discretionary presentation, workflow and strategy, never relevance, applicability, safety, authorization, privacy, required skill steps, validation or review. Without a valid principal session, personalization covers only organization settings (organization_scope_only) and the status says so; no status requires authentication or adds a prompt. A large profile arrives in pages: call again with the page_token the status names, plus the same keys and task_preferences when the status asks for them, until it says complete; a restart means the profile changed and earlier pages are outdated. Send known_digest only when this conversation already holds the complete snapshot for this exact context with that digest, every page included, with no summary or context reset since, and never for an incomplete, stale or cached copy; an unchanged profile then returns one short line. include_fingerprints adds each entry's fingerprint for compatibility feedback. Hosts that do not inject the snapshot each turn call this at each request start unless that request's query_skills or invoke_skill response already carries a complete snapshot or one saying no recovery call exists. Follow the part of a preference written for this host. Unlike an outage, one naming a model, tool, or agent this host lacks, with no part for this host, was written for another agent family: apply the closest equivalent this host supports and say which. When none is safe to assume, ask the user; if you cannot ask, keep your default and say so.
get_private_lesson_policy1 of 3 parameters describedFetch and cryptographically verify the authenticated organization's signed private-lesson automation policy. The server response never exposes an organization id and public_fallback is always false. The local MCP augments this with secure outbox readiness; hosted-only clients report auto_capture_supported=false.
record_preference0 of 9 parameters describedRecord a bounded, privacy-safe preference observation for the current verified installation. Built-in settings need only key and value; a custom setting also supplies its plain-language label, behavior, presentation/workflow/strategy_selection effect, prefer/avoid direction, and definition version. Send normalized settings and hashes only, never prompt text. A valid principal session is required. An explicit instruction governs the current task immediately. This call does not guarantee activation: read activated and normalized_activation_state in the response rather than assuming the setting is live. Known built-ins normally activate durably at once, custom settings remain pending until automatic normalization and validation approves them, and inferred observations activate only after Remembrance's consistency and confidence thresholds are met. Organizations on customer-held encryption are deferred: every substantive preference write is refused before persistence with HTTP 422 and code customer_held_preference_processing_deferred, existing records are unchanged, and reads plus the content-free withdrawal lane keep working. Do not retry that refusal, and do not offer preference compatibility feedback there. To keep an adaptation of another agent family's preference, record it for this runtime only: scope member_runtime, the same key, and a value, label, and behavior naming your equivalent; use source_category explicit_user only when the user confirmed it. Never overwrite the original in a shared installation or member scope. If member_runtime is refused because this installation is not linked to a member, keep the adaptation to this task and say once that linking the installation lets it persist.
submit_preference_compatibility_feedback0 of 9 parameters describedAfter actually opening or invoking a skill, record whether that exact version supports or conflicts with one active working preference, named by an offered preference_fingerprint or by its preference_key. An unavailable host tool or model is not by itself a skill conflict; runtime capability is a separate question. This is private to the authenticated organization, requires the same verified installation principal that used the correlated query/result, and never changes the skill or another organization's ranking. Use evidence_source explicit_user only for a user's explicit assessment; otherwise use agent_observed. Send only correlation IDs, the offered fingerprint or the setting key, exact slug/version id, relationship, source, and an optional allowlisted reason_code; never send prompts, source code, paths, free-text feedback, or a caller-generated idempotency hash. A key works for any preference that was active for this skill use, including one the offer did not list; a key that matches none, or more than one, is refused. Organizations on customer-held encryption are deferred: the server withholds the preference_compatibility_feedback offer there and refuses this call before persistence with HTTP 422 and code customer_held_preference_processing_deferred. Never retry a refusal.
link_current_installation0 of 1 parameter describedLink this verified installation to the signed-in organization member using a single-use dashboard token. The token expires after ten minutes and is consumed exactly once. A valid organization principal session is required; this never changes the organization API key or creates another billed agent.
get_skill3 of 3 parameters describedFetch the exact active reviewed skill body and version after query_skills returns a known slug. The response includes skill_md, version_id, and source. Pass query_id and result_id from that response so Remembrance can measure whether surfaced guidance was opened. Do not guess private or inactive slugs.
get_resource3 of 3 parameters describedFetch a known resource by slug after query_skills returns it. Pass query_id and result_id from that response so Remembrance can measure whether surfaced guidance was opened. Do not use for arbitrary URL fetching.
bootstrap_agent_identity4 of 5 parameters describedCreate or reuse a local Ed25519 TOFU key and register it with Remembrance. No arguments are required. Rerun to repair permissions or re-bootstrap if the local agent-key.json was lost.
acknowledge_feedback_denial0 of 4 parameters describedAcknowledge one exact submit_query_feedback or submit_feedback event that the host denied before transport. This writes only a private, content-free, hashed local marker with 14-day retention. It performs no network request, grants no approval, retries nothing, and never records feedback payload or draft content. Use result_id only for submit_feedback.
decline_pending_action2 of 2 parameters describedDecline one pending Remembrance action that a hook reminder listed and you cannot complete. Pass the action_id shown under it and one reason: not_observed_by_this_agent (you never received the result or payload it refers to), agent_exited (the subagent that earned it has exited), not_reusable (it holds nothing reusable), or duplicate (it was already completed another way); host_policy_denied only for a private-learning line or notice whose exact action the host denied. It is local-only: it sends nothing, stores nothing, grants no approval, and never counts as feedback. The local completion hook closes that action in your own pending list when it observes this call; a declined private draft stays retained: a lesson draft declined with any reason but host_policy_denied is sent by default after its first wait, by a later main-thread session that uses the API base and organization key it was captured for, and any other declined draft returns later only as a user decision.
get_pending_action_payload2 of 2 parameters describedReturn the exact payload a pending Remembrance action keeps on this machine, when a hook reminder names it with an action_id and a ledger instead of showing it. Pass both as shown. It reads only that ledger, verifies the payload against its recorded digest, and returns it unchanged with its kind and provenance: a reusable-evidence follow-up's submit_remembrance payload as payload, or a durable preference's exact record_preference arguments as arguments. Provenance is first_hand; second_hand when an exited subagent left it; subagent_identity_unavailable or pre_agent_keying when the record cannot say which agent earned it. It is local-only: it sends nothing, submits or records nothing, applies no preference, and changes nothing. A kept payload stays until its action is completed, declined or withdrawn. A payload that was never kept, or cannot be opened or verified, is reported as such and never guessed; then use only a payload or arguments you still hold verbatim, or decline the action. A missing key is recovered only by restoring the original key; a new key cannot open it.
capture_learning2 of 6 parameters describedThe zero-network routine reusable-learning entry point for a capture-capable local client; use it silently without a separate conversational permission step. Supply exactly one compact lesson, complete private skill, or structured preference observation. The local router chooses the organization-private lane, encrypts lesson and skill drafts before persistence, and returns one exact next action. The preference route persists nothing locally (persisted_locally false); record_preference completes it, so read that response rather than assuming the preference is live. Call that action immediately; the host asks for approval only when the private network lane is not already approved. Complete skills receive a persisted trusted inspector digest and exactly one outcome: unchanged, redacted (source-redacted), positive_recovery_need, or operational_failure. Only a positive recovery need for confirmed source-ownership ambiguity or a recognized encrypted private container returns agent_recovery. Capacity, internal, and inspector-contract faults are operational failures, never semantic ambiguity or permission to reformulate. A replacement_of submission must be a fresh complete skill, preserve adequate task meaning, and pass full inspection; recovery context is separately inspected, explicitly partial, and never the unchecked held original. Each inspector/context epoch accepts at most three distinct candidate digests; only a new authenticated native-user session context or trusted inspector identity opens another. The original and attempts remain encrypted until private acceptance, which completes the draft through the existing private_skill approval lane without another approval. If adequate context is unavailable, retain the draft; do not invent instructions or ask the user to inspect files. It never accepts organization, destination, visibility, credential, or public-fallback fields. Query-result fit is intentionally excluded: always use submit_query_feedback for good, partial, or poor query matches.
inspect_private_learning_outbox1 of 2 parameters describedInspect local state for both encrypted private-learning outboxes without a network request. Listings contain per-lane status, draft IDs, retry metadata, retained bytes, trusted persisted inspector outcomes/digests, and recovery reason codes without original candidate payloads or paths. Set private_skill_draft_id to inspect one exact draft independently of pagination. The four skill outcomes are unchanged, redacted (source-redacted), positive_recovery_need, and operational_failure. Only positive recovery need can expose separately inspected, explicitly partial native context and agent_recovery; capacity and other inspection faults remain operational. The held original is never returned as context. Recovery uses at most three distinct candidate digests per inspector/context epoch. Candidate text, duplicates, and timers do not reset the bound; only new authenticated native-user session context or trusted inspector identity can open another epoch. Follow eligible recovery autonomously through the existing private_skill approval lane. Unavailable context and operational failures remain encrypted for a later valid lifecycle opportunity, not submitted or discarded; do not invent instructions or ask the user to inspect files. One corrupt lane cannot hide the other.
queue_private_skill_import0 of 4 parameters describedLocal-only fallback for an organization-approved skill contribution that a host privacy or network policy would not send. Writes a mode-0600 organization-private handoff bundle without contacting Remembrance. It never submits or syncs the content: an organization admin must deliberately upload the file through Dashboard > Skills > Import, where every skill enters the normal private review flow. Do not use another transport to evade a host denial and do not claim the skills were submitted.
prepare_private_lesson_candidate2 of 6 parameters describedLocally canonicalize, redact, validate, encrypt, and durably queue one compact organization-private failure, correction, or evidence lesson. Limits apply after local redaction: generalized_lesson at most 1200 characters and each stable_conditions item at most 240; an over-limit call is refused with the field and its length, and nothing is stored. This is an evidence lane, not a complete-skill transport: use propose_private_skill or propose_skill_idea for a complete reusable procedure, playbook, or actionable instructions. Tags use an open, bounded lowercase slug vocabulary; a malformed or privacy-sensitive tag is dropped and counted. Every redaction match becomes its [redacted:category] marker and the draft continues; only a stack trace or zero-width or bidirectional control characters hold it as held_safety. Pre-redaction input remains memory-only and this tool performs no network request. The result returns a draft handle and never echoes lesson prose. Follow the returned next_action with submit_private_lesson_candidate; held_safety content remains local. An unsupported redactor version or digest becomes terminal superseded_redactor and is never retried or re-redacted.
inspect_private_lesson_outbox2 of 4 parameters describedInspect content-free state, retained bytes, terminal reasons, deletion guidance, retry metadata, and one reachable page of encrypted local private-lesson drafts. Use offset with next_offset until truncated is false; selecting draft_id remains independent of pagination. This never returns lesson prose or canonical payload bytes. Terminal drafts remain retained and count toward the safety ceiling until explicitly deleted.
submit_private_lesson_candidate1 of 10 parameters describedSubmit exactly one canonical, challenge-attested organization-private lesson candidate or one content-free held-safety telemetry event. Local plugin clients normally pass only the draft_id returned by capture_learning; raw or hosted clients pass the strict candidate-or-hold union. The action always creates only an organization-private submission and never falls back to anonymous or public submission. Held telemetry contains no lesson prose and can never enter verification, review, topology, or skill materialization. Exact replay is idempotent; host denial leaves the local draft untouched.
retry_private_lesson_candidate3 of 3 parameters describedRetry one retained private-lesson draft through the same private_lesson approval lane as submit_private_lesson_candidate. Network access is allowed only for authorization-pending, due transient retry, newly compatible profile, or refreshable policy states, and for a held draft the current local safety check releases once the user chose to deliver it (deliver_rechecked_hold=true); a fresh ready draft is rejected locally. A legacy_destination_recovery review explicitly requests a content-free signed-policy review for one eligible unbound draft, without changing or sending it. Only new explicit owner approval of that exact review and all outbox readers at 0.2.6 or newer permits local destination binding; a separate normal retry performs delivery. Never infer migration approval from submission approval. It never falls back to a public or alternate transport.
delete_private_lesson_candidate2 of 2 parameters describedExplicitly delete one selected local private-lesson draft. confirm=true is required. No automatic cleanup exists, and deletion leaves only a content-free tombstone.
submit_private_skill_candidate1 of 1 parameter describedSubmit exactly one encrypted local complete-skill draft to the authenticated organization's private review and verification pipeline. Pass only the draft_id returned by capture_learning. The action never accepts a visibility or organization override and never falls back to public submission. An idempotent private acceptance completes the selected original or replacement and removes the encrypted original, recovery contexts, and attempt content; recovery reuses this existing private_skill approval lane rather than requesting another approval.
retry_private_skill_candidate1 of 1 parameter describedRetry one retained private-skill draft through the same private_skill approval lane as submit_private_skill_candidate. Network access is allowed only for authorization-pending or due transient retry states, an explicit same-draft retry after an administrator corrected a private-protection configuration hold, and for a draft an older server refused only for length: that waits out any delay the server set, then reads its signed capability and resends the exact draft only if it admits it. Configuration holds never schedule automatic retries; an unchanged refusal remains held. A fresh ready draft is rejected locally. It never switches transport, visibility, key ownership, or protection mode.
delete_private_skill_candidate2 of 2 parameters describedExplicitly delete one selected local private-skill draft. confirm=true is required. Unsubmitted drafts are never removed automatically; successful content is removed immediately and its content-free completion marker expires after 14 days.
submit_feedback1 of 10 parameters describedSubmit categorical post-use skill feedback without a separate conversational permission request: useful or not useful, an optional rating, and the originating query_id/result_id when the skill came from query_skills. Reusable prose is deliberately not accepted here; call capture_learning separately for a generalized lesson, complete private skill, or preference. This tool never replaces submit_query_feedback, which remains the only query-fit verdict path.
submit_query_feedback2 of 7 parameters describedSubmit one complete set of bounded explicit good, partial, or poor judgments for result_id values from one query_skills call, using the same auth scope as that query. Use overall_fit none only when every rated result is poor; if any result is partial or good, use partial or good as appropriate. Keep honest per-result judgments when correcting validation errors; do not drop them or relabel matches to bypass validation. Each result_id appears once; poor ratings require at least one reason. This bounded feedback action carries no reusable content and needs no separate conversational permission request. Unrated results stay neutral. Notes are optional and compact; this path remains independent of private learning capture and never accepts a reusable lesson. Keep routine organization ratings minimal: use the Remembrance-issued query_id and result_id values, judgments, required reasons, and an idempotency key when needed, and do not automatically attach internal artifact hashes, task text, source snippets, or raw logs. An identical retry is safe, but changed later judgments conflict. Labels adjust bounded retrieval profiles without changing global skill usefulness; use submit_feedback only after actually using a skill.
report_task_outcome0 of 15 parameters describedReport completion or abandonment for a query. Remembrance accepts one terminal outcome per query; retry the same report with the same idempotency_key, and do not submit a different later outcome. Include only result_ids listed in task_outcome.eligible_result_ids; each result and bundle also carries task_outcome_eligible. When two or three selected skills exactly match a returned skill_bundles entry, include its bundle_id for value attribution; other combinations are accepted as funnel telemetry but do not train a skill or bundle cohort. Include bounded token totals only when the runtime exposes them. For Vercel AI Gateway work, include metering_reference.adapter=vercel_ai_gateway and every gen_ generation ID used by the task (maximum eight); Remembrance retrieves authoritative usage asynchronously, and caller totals never establish proof trust. Never send prompts, outputs, source paths, private URLs, transcripts, or proprietary task content. Completion without token counts still closes the funnel but cannot establish metered savings proof.
get_value_proof1 of 1 parameter describedFetch and cryptographically verify the token-only signed proof referenced by a qualified potential_savings estimate. Successful results include signature_verified=true. Public-skill proofs are anonymous reads; organization-skill proofs require an active query-capable API key from the same organization, but not necessarily the key used for the original query. Proof payloads contain no organization identity, task identity, price, credit, subscription, or payment data.
submit_remembrance1 of 14 parameters describedSubmit a full remembrance payload with detailed reusable evidence, including high-value self-corrections, user-caught mistakes, CI/deploy failures, and release/versioning misses as type failure_report. Never include credentials, secrets, raw logs, or unrelated repository content. Repository-derived private instructions may be sent only when the user or organization has approved Remembrance as a destination. If the host blocks that export, report that nothing was sent and do not retry through another transport or create a handoff unless an organization admin explicitly requests one.
propose_skill_idea4 of 12 parameters describedPropose a complete missing reusable skill when query_skills has no useful result. This full skill lane retains bounded markdown, public citations, snippets, metadata, and actionable instructions through normal safety and review; do not compress a finished skill into the narrow private-lesson contract. THIS TOOL'S VISIBILITY IS NOT FIXED: an active organization API key keeps the candidate inside that organization's review queue, while intentionally calling without a key creates a PUBLIC candidate. A supplied invalid/inactive key fails with 401 and a valid key without submission:create fails with 403; neither failure creates a candidate. Prefer propose_private_skill for anything repository-derived, organization-specific, or otherwise not intended for the public registry because that endpoint structurally requires organization auth. Use propose_skill_idea only when a public candidate is an acceptable outcome. Check `visibility` (organization_private or public_candidate) and `owner_scope` in the successful response, and report where the candidate landed. Never remove or bypass an organization key to force public submission: submit privately, then use the reviewed public-propagation flow. Send repository-derived private instructions only when Remembrance is an organization-approved destination. If host policy blocks the export, report that nothing was sent, do not retry, and do not create a handoff unless an organization admin explicitly requests one. A complete skill also needs proposed_metadata.applicability and .routing (routing contract v1; see that property): set prefill_routing_metadata to get a platform draft first, show it to the author, then resend the confirmed or edited fields with routing_draft_token. A 422 skill_routing_metadata_incomplete stores nothing and carries the same draft.
propose_private_skill5 of 12 parameters describedSubmit a complete proposed skill only to the authenticated organization's private review queue. This full skill lane retains bounded markdown, public citations, snippets, metadata, and actionable instructions through normal safety and review; do not compress a finished skill into the narrow private-lesson contract. proposed_skill_md accepts up to 56000 characters, as organization skill imports do, and the whole JSON request must fit the organization's request-body limit (262144 bytes by default); a larger request is refused whole and creates nothing, never truncated or split. Privacy here is STRUCTURAL, not credential-dependent: this endpoint rejects anonymous use and never creates a public candidate, so unlike propose_skill_idea it cannot silently fall through to public when a key fails to resolve — it fails closed with 401/403 instead. Prefer it whenever the content is repository-derived, organization-specific, or must not reach the public registry; that preference is about removing the credential from the privacy decision, not about propose_skill_idea being unsafe when a key is present. Use it for repository-derived instructions only when Remembrance is an organization-approved destination. The response reports `visibility` (always organization_private here) — a rejection means nothing was submitted. If host policy blocks the export, report that nothing was sent, never retry through another transport, and do not create a handoff unless an organization admin explicitly requests one. A complete skill also needs proposed_metadata.applicability and .routing (routing contract v1; see that property): set prefill_routing_metadata to get a platform draft first, show it to the author, then resend the confirmed or edited fields with routing_draft_token. A 422 skill_routing_metadata_incomplete stores nothing and carries the same draft.
submit_suggestion0 of 11 parameters describedSuggest a safe update to an existing skill. Do not submit prompt-injection text, secrets, or install URL changes.
submit_resource0 of 4 parameters describedSubmit a discovered reusable API, MPP endpoint, MCP server, docs site, package, dataset, service, or tool.
submit_resource_review0 of 6 parameters describedSubmit a review after using a resource. Include ratings and redacted evidence only.
request_attestation_challenge0 of 9 parameters describedRequest a challenge for manually signing remembrance or resource-review evidence.
register_agent_key0 of 8 parameters describedRegister a TOFU public key with a proof signature. Prefer bootstrap_agent_identity unless you manage keys yourself.
Tested in a sandbox with an empty environment: no API keys, a throwaway folder, nothing from a real account. A passing test shows the server runs and responds. It is not a security audit.