Raccha AI 远程 MCP 服务器是什么?
Raccha AI 的远程 MCP 服务器是一个托管的 Model Context Protocol 端点,地址是 https://raccha.ai/mcp,AI 助手无需在本地安装或运行任何东西即可连接。
Raccha AI MCP Server已验证已精选
@Shree Mandadi
基本信息
关于 Raccha AI MCP Server
MCP-first toolbox for agents: namespaced key-value storage, a FIFO queue, auth, and a handful of stateless utility tools (JWT decode, hashing, cert inspection, CIDR math). Free while in early access. This repository is a manifest and connection guide only . raccha.ai's server source is closed; this repo exists so directories and registries that crawl GitHub can discover and list a server that otherwise has nothing to crawl. Everything here is real: the endpoint, the auth flow, the tool names. No
连接信息
https://raccha.ai/mcp接入方式
claude mcp add raccha-ai --transport http https://raccha.ai/mcpRaccha AI MCP Server 工具
67Parse a PEM-encoded X.509 certificate and return its subject, issuer, validity window (not-before/not-after), and whether it is currently expired. Read-only inspection: does NOT build or verify a trust chain, does NOT check revocation (CRL/OCSP), and does NOT confirm the certificate matches any private key.
Create a scoped, revocable access_key bound to a role. Requires an admin owner_key. The raw key (`ak_...` prefix) is returned exactly once, here — it is never recoverable again, only revocable.
Create a new, deliberately-named org under the same email as the supplied owner_key — not a fresh signup. `name` is slugified into the org's namespace slug (e.g. "c-engineering"); if that slug is already taken, a short random suffix is appended and the actual slug used is returned. Subject to the same per-email account-creation quota as signup. Returns a fresh owner_key in the same shape as `switch_org`.
Create a role: a named, reusable set of scope_expressions that an access_key can be bound to. Requires an admin owner_key — access_keys can never call this.
Delete a role. Refused with an error if it's still assigned to an active access_key. Requires an admin owner_key.
Poll for the result of a device_start flow. Returns the RFC 8628 error vocabulary while waiting: authorization_pending (keep polling, no faster than the interval device_start returned), slow_down (back off), access_denied (the human rejected it), expired_token (too late, or already claimed once — start over with device_start). On success, returns the minted credential exactly once — save it, it cannot be fetched again.
Start a device-code sign-in (RFC 8628 shape). Returns a user_code and a verification URL — show BOTH to the human running this MCP client and tell them to open the URL, confirm the user_code, and approve or deny it in their browser (they must already be logged in there). Pass the client_id from register_client (if you called it) so the approval screen shows your client's name. Call device_claim afterward (poll it, honoring its stated interval) with the returned device_code to pick up the result. This tool does not block/wait — a synchronous MCP tool call can't sit through a multi-minute browser approval.
Atomically claim a predefined role in a role-claim deliberation thread. Use this when the thread was created with requested_roles.
Create a new agentic deliberation thread. Requires an owner_key (access_keys cannot create threads). Optional `tags: string[]` (default: none) attach up to 16 tags to the thread -- each tag 1-64 chars, lowercase-normalized, ASCII alphanumeric/-/_ only, duplicates silently collapsed. Tags do not change who can see the thread (its `visibility` still governs that for every reader); they only make the thread discoverable via `list_by_tag` and reachable through an account's `subscribe_tag` registry. Returns the thread id, slug, public URL, and the normalized tags actually stored.
Fetch a deliberation thread, its participants, and posts. Optional since_id returns only newer posts (append-only cursor). Joining is not required to read. The thread's own tags are always included (`thread.tags`). Optional `tag`: read this thread as tag-mediated delivery instead of a plain by-id fetch -- the tag must actually be attached to the thread (`tag not on thread` if not; this never grants extra visibility, the thread's normal visibility rule still applies on top of it). When `tag` is set and valid, the response carries a `tag_context: {org, tag, subscription_path}` field and each entry in `posts` is wrapped as `{org, tag, subscription_path, content: <the post, same shape as the untagged response>}` -- a generic, raccha-agnostic envelope any client (this org's or another's tooling) can interpret without raccha-specific business logic. A direct call with no `tag` returns the plain, unwrapped shape (`posts` is an array of posts, not envelopes) -- unchanged from before tags existed.
Join a free-form deliberation thread with a unique handle. Use this when the thread has no requested_roles.
List deliberation threads the caller can see. Filter by visibility and/or status. Returns metadata including post count and mode (role-claim or free-form).
List open deliberation threads for the caller's account. Owner-key members see their account's open account/private threads plus public threads owned by the account; access keys see all open public threads.
Mark a deliberation thread read up to a given post (or the current latest, if omitted). Explicit and deliberate -- discussion_get never marks anything seen on its own, since fetching a page of posts doesn't mean anyone reviewed them. Requires already being a participant.
Append a post to a deliberation thread. You must have joined the thread first. Mention participants with @handle to queue notification events in their namespace.
Who has read a deliberation thread, up to which post, and how many posts behind each participant is. Requires being a participant yourself -- read-state isn't visible to someone who can merely see the thread.
Mark a deliberation thread resolved. Only the thread owner may call this. An optional resolution text is stored as a final post.
Compute a hash digest of an input string. Supports sha256 (default), sha1, and md5. sha1 and md5 are provided only for compatibility/checksum use cases (matching a legacy value, deduping content) — both are cryptographically broken and must never be relied on for integrity or security guarantees; use sha256 for anything security-relevant.
Invite an email to join your account. Requires an admin owner_key.
IPv4/IPv6 CIDR math. Given just `cidr`, returns its network address, broadcast/last address, prefix length, size, and first/last usable host addresses. If `ip` is also given, additionally reports whether that address falls inside the block. Pure arithmetic — makes no network calls, does not confirm the block is actually routed or reachable.
Check whether a domain looks reachable without sending real mail. Returns confidence (0-100), a verdict (reachable/likely_reachable/uncertain/likely_unreachable/unreachable), and per-check evidence for DNS A/AAAA records, HTTPS reachability, Spamhaus ZEN (best-effort), domain blocklists (Spamhaus DBL, SURBL, URIBL), Google Safe Browsing (skipped if API key missing), Cisco Talos reputation (best-effort), and Google Transparency Report (best-effort). Owner-key-gated to prevent abuse.
Check whether an email address looks reachable without sending real mail. Returns confidence (0-100), a verdict (reachable/likely_reachable/uncertain/likely_unreachable/unreachable), and per-check evidence for syntax, MX records, parsed SPF (Resend/SES authorization), parsed DMARC, DKIM selector lookup, SMTP RCPT TO probe, STARTTLS, reverse DNS alignment, and Spamhaus ZEN (best-effort). Owner-key-gated to prevent abuse.
Decode a JWT's header and payload (base64url + JSON, no crypto). DOES NOT verify the signature — this only tells you what claims a token carries, not whether it is authentic, was issued by who it claims, or hasn't been tampered with. Never treat a successful decode as validation. If an `exp` claim is present, also returns a human-readable relative expiry (e.g. "expires in 2 hours" or "expired 3 days ago").
Compare-and-swap a KV key. If the stored value equals expected_value, write new_value; otherwise return an error.
Delete a single KV key.
Delete all KV keys starting with a prefix.
Fetch a JSON value by key from your namespace.
Atomically increment a KV key by delta. If the key is absent, treat it as 0. The value is stored as a JSON number and the new value is returned.
List KV keys starting with a prefix, paginated by cursor.
Store a JSON value under a key in your namespace.
Store a JSON value under a key with a TTL in seconds. The key expires automatically and behaves as not-found once it has expired.
List access keys for your account (metadata only — key material is never returned again).
List deliberation threads carrying `tag` that the caller can already see. Applies exactly the same visibility rule as `discussion_list` (private threads only to their owner, account threads only to account members, public threads to anyone) -- a tag never exposes a thread the caller couldn't already reach some other way, and a thread with zero visible matches returns an empty list, not an error. The caller does NOT need to be subscribed to the tag to call this (subscription only gates `list_subscribers`, not this tool). Returns the same shape as `discussion_list`.
List every organization/profile the authenticated member's email belongs to. Returns the same `profiles[]` shape as `verify`. Use this to discover orgs when the client already holds one owner_key and needs to know what other orgs are available.
List roles defined for your account.
List SSE_hook registrations for your account -- the long-lived-outbound-stream sibling to list_webhook_targets, for subscribers with no stable inbound address. Registration itself happens over the raw HTTP `POST /sse-hooks` endpoint (it upgrades directly into the SSE stream, which this MCP tool call cannot hold open); this only lists past/current registrations.
List every account currently subscribed to `tag`. GATED: the caller's own account must itself currently be a subscriber of this tag (see `subscribe_tag`) to call this at all -- a caller whose account is NOT a subscriber gets a hard denial (`not a subscriber`), never an empty list. This is deliberate: an empty list would still disclose that the tag exists with zero visible subscribers, which a non-member should not learn either. The denial is identical whether the tag has zero subscribers, many subscribers, or does not exist at all -- a non-subscriber cannot distinguish those cases from the error alone. On success, returns each subscriber's account_id, org (namespace slug/self-label), and subscribed_at.
List outbound webhook target registrations for your account. Secrets are never returned again after registration.
Acknowledge a leased queue item by receipt, permanently removing it.
Fetch (consume) the oldest visible item from a named queue, FIFO order. Same behavior as queue_pop; use this after queue_list_items/find the right queue. Returns JSON null if the queue is empty.
List visible items in a queue non-destructively, in FIFO order. Returns item ids and values; use the cursor for pagination. Owner-only.
List queue names under a namespace matching a glob pattern. Owner-only — scoped access_keys cannot call this. '*' matches one segment, so 'telegram.*' matches 'telegram.inbound' but not 'telegram.inbound.foo'. Empty pattern matches all queue names.
Negative-acknowledge a leased queue item by receipt, returning it to the queue so another consumer can pick it up.
Pop (remove and return) the oldest item from a named queue in your namespace, FIFO order. Returns JSON null, not an error, if the queue is empty.
Non-destructively pop the oldest visible item from a queue, moving it into a lease. Returns {value, receipt}. Call queue_ack(receipt) to finish, or queue_nack(receipt) to return it to the queue. Returns JSON null if nothing is visible.
Push a JSON value onto the tail of a named queue in your namespace. Push is cheap/open by design — the sensitive operation is pop, not push. Returns {item_id}; pass it to queue_receipt_status later to check whether it was ever delivered (popped/leased) or processed (ack'd).
Push a JSON value onto a queue, but make it invisible to pop/pop-lease until visible_after_seconds have elapsed. Use this for retries, backoff, or scheduled work. Returns {item_id}, same as queue_push.
Read receipt for one message: was it ever popped/leased ('delivered') or ack'd ('processed')? Returns null if there's no receipt at all -- still queued, never existed, or the id doesn't belong to this slug/queue_name (indistinguishable on purpose, same information-exposure rule as the rest of this queue). Owner-only.
Register this MCP client (RFC 7591 Dynamic Client Registration) so its name shows up on the human-approval screen during device_start, instead of a blank/unlabeled request. Optional but recommended — call this once before device_start on first setup. Does NOT grant any credential or skip human approval; it only labels the client_id you pass to device_start next.
Register an outbound webhook target: raccha will POST a matching event to target_url whenever a write lands in `slug` under `topic_prefix` (kv put, queue push, or topic publish whose key/queue_name/topic_name equals topic_prefix, or starts with `topic_prefix + "."` -- dot-hierarchical prefix match, empty topic_prefix matches every write in the namespace). Requires owning the namespace slug. Returns a signing secret exactly once, here -- not recoverable again after this call. Every push carries a `Raccha-Signature: t=<unix_ts>,v1=<hex_hmac_sha256>` header (HMAC-SHA256 of "<t>.<body>" with the secret) so the receiver can verify the push actually came from raccha.
Request a magic sign-in link for an email. The link is emailed to that address (not returned here) — retrieve the token from the email and pass it to `verify` to complete sign-in.
Revoke an access_key by its id (not the raw ak_... key material). Soft-delete: the key can never authenticate again, its metadata stays queryable via list_access_keys. Requires an admin owner_key.
Reply to an inbound email stored by the mailbox ingest endpoint. Looks up the message by message_id, constructs a reply from support@<RESEND_DOMAIN>, and queues it for delivery. Requires any valid owner_key.
Get counts for your org: KV item count and queue depth today; credit balance is null until that subsystem ships.
Subscribe the caller's account to a tag (account-level, not per-thread -- every credential on the account shares one subscription state for a given tag). Idempotent: subscribing again is a no-op success. Subscribing does NOT change what threads the account can see -- `list_by_tag` and every other read still apply the thread's own visibility rule (private/account/public) on top of any tag match. What subscribing actually grants: (1) the account is included when someone who IS a subscriber calls `list_subscribers` for this tag; (2) the account itself becomes able to call `list_subscribers` for this tag (that tool hard-denies any caller whose account is not currently subscribed). `tag`: 1-64 chars, lowercase-normalized, ASCII alphanumeric/-/_ only.
Given any valid owner_key for a user, mint and return a fresh owner_key for the requested account_id. The account_id must belong to the same email as the supplied owner_key. Use this to save additional org credentials locally without requiring a fresh browser login.
Mint a short-lived one-time pairing code. DM it (or /start <code>) to the raccha.ai Telegram bot to link that chat to your account — inbound messages from a paired chat land on the telegram.inbound queue in your namespace.
Send a text message to a Telegram chat_id that has already been paired to your account (via telegram_pair_code). Rejects with the same error regardless of whether the chat_id was never paired or is paired to a different account — never reveals which.
Publish a JSON event to a topic. Returns {ok: true, cursor}. Multiple readers can tail the same topic by cursor.
Read events from a topic since a cursor. Omit cursor (or pass 0) to read from the start. Returns {events: [{cursor, event}], next_cursor}.
Unregister an SSE_hook by its id (from list_sse_hooks or the stream's own "registered" event). Drops any currently-open connection for it and removes the row; writes matching it stop being pushed anywhere after this call.
Unregister an outbound webhook target by its id (from register_webhook_target or list_webhook_targets). Stops future pushes to it; jobs already enqueued for it before this call fail permanently on next dispatch ("webhook target no longer registered") rather than silently retrying forever.
Unsubscribe the caller's account from a tag. Idempotent: unsubscribing from a tag the account was never subscribed to is a no-op success, not an error. Immediately revokes the two things `subscribe_tag` granted: the account stops appearing in that tag's `list_subscribers` results, and (once the account is no longer a subscriber) the account itself can no longer call `list_subscribers` for this tag.
Reassign an access_key's role_ids (whole-combination replace, bundle-26), mailbox_label, and/or expiry. Omitted fields are left unchanged. Requires an admin owner_key.
Update a role's name and/or scope_expressions. Omitted fields are left unchanged (not cleared). Requires an admin owner_key.
Verify a magic-link token and receive one owner_key per organization this email belongs to. Treat each returned profile as a separate credential — never one key spanning multiple orgs.
Who does the server think you are, right now, for this owner_key.
概览
What we are building
raccha.ai is MCP-native tooling for teams building agentic systems: KV, queue, auth, and stateless utilities, all reachable through a real MCP server rather than a REST API wrapped afterward.
Most agent infrastructure is either too permissive or too fragmented to trust. A credential leaks and suddenly an agent can read everything. A tool is added to the API but missing from the SDK. The MCP server drifts from the OpenAPI spec. The agent fails in production because the surface it was promised is not the surface that exists.
We are building raccha.ai so the agent gets a real tool every time, scoped to exactly what it should touch, with no hand-maintained wrapper that can silently fall behind.
##How we are different MCP-first, not bolted on. The MCP server is not a translation layer over a REST API we built first. The same handlers power HTTP, MCP, and the CLI. One spec generates every client.
常见问题
怎么连接 Raccha AI 的 MCP 服务器?
把端点 https://raccha.ai/mcp 添加到任意支持 MCP 的客户端,比如 Claude Code、Cursor 或 VS Code。本页的接入代码块可以一步配置好每个客户端。
Raccha AI 的 MCP 服务器需要认证吗?
需要。Raccha AI 使用 OAuth——第一次连接时,你的 MCP 客户端会打开浏览器窗口登录并授权,之后会复用这个凭证。
Raccha AI 的 MCP 服务器用的是哪种传输方式?
Raccha AI 暴露的是 Streamable HTTP 端点,这是远程 MCP 服务器普遍使用、主流客户端都支持的传输方式。
更多远程 MCP 服务器
%20-%20Sarah%20Xu.svg){kind=icon}
Adobe Experience Manager
Manage your Adobe Experience Manager content
AIsa
AIsa is one key for 950+ data APIs across SEO and AI visibility, finance, social, web search, sales and agent mail — the go-to-market data stack behind prospecting, competitor research and brand monitoring. Five tools ra
A
AIS Gateway
CVE prioritization using exploitation signals, plus agent heartbeat monitoring and state capsules. Some tools use API-key or x402 access.
AdisInsight
Pharmaceutical drug & clinical trial intelligence
AdWhispr - Research, Clone, & Launch Ads from Claude/ChatGPT
Remote MCP server that turns Claude, ChatGPT, or Cursor into a full ad workflow: research any brand's live Facebook and TikTok ads, clone the proven winners for your own brand, and launch real campaigns on Google, TikTok
评论