omniseek
OmniSeek
你的智能体寻找搜索找不到的东西。
答案藏在播客的第47分钟、评论串下三层的回复里、登录墙后面,或者另一种语言中。你的智能体照样能拿到。
自托管感知 MCP 服务器 · 一个连接
搜索能给智能体的只有被索引的网页,一种语言,纯文本,然后到此为止。
OmniSeek 则会继续搜寻:跨越语言、登录墙、评论串、音频和像素,而且全部在你的机器上完成。
它能听(本地双语自动语音识别,无云端)、能看(图像与视频帧,带内处理)、能跨语言(中文查询找到英文结果,反之亦然)、能读登录墙后面的内容(你的凭证,你的机器,默认关闭)、能记忆(持久化检索记忆,加上有类型、可溯源到来源的证据图谱)。
跨语言能力依赖于 OmniSeek 在使用过程中构建的索引,因此全新安装从基线开始。已发布的声明验证测试正是跑在全新安装上,所以它们的跨语言数值是最冷的情形,而不是典型情形。
目录中的每个源,即 OmniSeek 能触达的一切的精选名册,都是因为在五类任务(结构化、破墙、转写、召回、监测)之一上胜过普通搜索才赢得一席之地:引用图谱、监管文件、登录墙论坛、中文视频。而且目录本身是可增长的:策展流水线会探测、评估并接纳新源,也会淘汰退化的源。
实际示例,真实输出 · 完整案例研究 · 以上每一项声明都是一项测试 (最新结果) · 来源健康状态,每周更新
快速开始
Docker(推荐)
git clone https://github.com/Battam1111/omniseek.git && cd omniseek
docker compose up -d
docker compose logs omniseek # bearer token printed on first start
curl -s http://127.0.0.1:8765/healthz将 MCP 客户端指向 http://127.0.0.1:8765/mcp,并使用 Authorization: Bearer <token>。令牌在首次启动时生成,存储于 ~/.omniseek/credentials/omniseek_http.json(在宿主机上使用 compose 文件时,路径为 ./.omniseek/credentials/omniseek_http.json)。
从这里出发有两条路径。预构建的核心镜像(docker pull ghcr.io/battam1111/omniseek,amd64 + arm64)无需构建即可使用,并包含所有核心感知能力;它遵循 Apache 许可,不附带 PDF 阅读、听觉(ASR + 视频帧)和登录墙来源。想要这些额外能力才会触发本地构建:设置 EXTRAS="[pdf,asr,walled]" 并运行 docker compose build,然后 up -d(首次构建还会获取 headless Chromium;之后的启动则是即时的)。可选但建议:设置 OMNISEEK_CONTACT_EMAIL,可在 Crossref、SEC 和 Unpaywall 上获得更快的通道。
不使用 Docker
python -m venv .venv && . .venv/bin/activate
scripts/bootstrap.sh
python -m omniseek.serve_http基础安装对应 Core 层级:所有无需密钥的 API 和静态来源、除 PDF 外的文档读取,以及词汇记忆索引。pip install "omniseek[pdf,asr,recall,ocr]" 唤醒 Research 层级(PDF、听力、跨语言向量、OCR);omniseek[walled] 增加登录墙层级,该层级在你提供自己的账号之前保持关闭;omniseek[all] 安装一切。服务器在每次启动时都会打印哪些感知在线、哪些处于休眠。
在 Windows 上,通过 Git Bash 或 WSL 运行 bootstrap.sh;Docker 是最简单的路径。对于需要常驻的 Linux 服务,请参阅 deploy/omniseek.service。
更喜欢 stdio?安装包还附带一个 omniseek 命令,通过 stdio 使用 MCP 协议进行通信,供自行启动服务器的客户端使用;Dockerfile.stdio 在容器中封装了同样的功能。
OmniSeek 绑定 127.0.0.1,并严格要求每个请求都携带 bearer 令牌。没有反向代理时请勿暴露(SECURITY.md)。
Related MCP server: Augent
工具
一个 MCP 连接;内部没有模型,也没有智能体循环。你的模型负责思考,你的主控框架驱动循环,OmniSeek 负责触达。从 omniseek_search 开始;用 omniseek_sources 探索可用内容。
工具 | 功能 |
| 在整个目录中扇出检索,去重,排序。跨语言(语义 + 词法)。 |
| 将任意 URL 或文档(网页、PDF、arXiv)规范化为干净文本。 |
| 使用视觉读取图像、文档图表、视频帧。 |
| 在本地转录音频/视频。双语 ASR,可按时间戳切片。 |
| 描绘研究领域的引文邻域:基础核心 vs. 前沿。 |
| 将人名解析为跨数据库的候选作者 ID。 |
| 根据合著论文数量描绘研究者的合作网络。 |
| 列出某实验室中活跃发表者,限定于特定领域。 |
| 论文的开放获取 PDF、撤稿/诚信状态、引用数。 |
| 语义相似的论文(SPECTER 嵌入),关键词搜索会漏掉的那些。 |
| 查询累积的证据图谱:查找、邻域、路径、时间范围、相似。 |
| 带新颖性检测的常驻查询。只告诉你什么是新的。 |
| 记录图谱在读取时应用的身份判定(相同/不相同)。 |
| 记录图谱向前携带的有向关系。 |
| 来源生命周期:提交、探测、评估、接纳、退役。 |
| 读取来源接纳队列或单个来源的审计档案。 |
| 并行运行多个工具,一次返回。 |
| 列出并路由:域名、地区、能力、健康状态。 |
登录墙层级没有自己的工具:一旦你按来源选择启用,同一个 omniseek_search(..., sources=["xiaohongshu"], raw=True) 就会通过你自己已登录的浏览器运行。参见 登录墙来源。
在使用 Claude Code?skills/omniseek-investigate 把调查方法论(sweep、zoom、structure)打包成一个现成的技能。
配置
OmniSeek 是 目录优先:无需配置即可让所有良性来源开启,登录墙来源关闭。只需在一个文件中调整:~/.omniseek/profile.json(示例):
层级 | 默认 |
免费(公开,无需密钥) | 开 |
带密钥(你提供免费或付费 API 密钥) | 设置密钥后即开 |
登录墙(你持有一个登录账号) | 关;需要你自带浏览器 |
规避 | 关;默认包中不包含 |
为什么选择自托管
没有 OmniSeek 云。没有遥测、没有账号、没有中继:查询离开你的机器时,只会成为发往你所启用来源的直接请求,OmniSeek 不会在这条路径上增加任何其他方。登录墙来源的凭证留在你自己的浏览器中,只提供给它们所属的站点;OmniSeek 从不存储、上传,甚至不查看你的密码。它数月积累的检索记忆和证据图谱,都是归你所有的本地文件:停止运行 OmniSeek,你仍然拥有一切。这不是一个功能开关,而是架构。
贡献
参见 CONTRIBUTING.md。新来源的门槛:它必须在某一种模式(结构化 / 破墙 / 转写 / 召回 / 监测)上胜过普通网页搜索。修复退化来源的门槛:很低,欢迎去做。在推送前运行 python tests/smoke.py。
参与即表示你同意 行为准则。
你的智能体寻找搜索找不到的东西。
Apache-2.0 · NOTICE · 安全 · 引用
Available Tools
18 toolsomniseek_coauthorsARead-only
Use WHEN you want WHO a researcher collaborates with: advisor + closest collaborators by joint-paper count, or how a paper's author group is connected (WebSearch cannot build this). One LAYER, not the whole graph: co-authorship is one edge type; the other layers, and what each tie means, are yours.
Tool name: omniseek_coauthors (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_coauthors".
Args: authors: names, ids (omniseek_resolve_identity) or "id1+id2" for one person split across ids. source: "openalex" (default) or "s2". hints: per-author disambiguation hints, parallel to authors. papers: per-author known papers, parallel to authors.
One author: their coauthors ranked by joint papers. Several: also the pairwise joint-paper edges among them and the bridge collaborators shared by two or more. Returns {source, n_authors, nodes: [{query, resolved, ambiguous, alternatives, works_seen, top_coauthors: [{id, name, joint}], degraded?}], edges: [{a, b, joint_count, papers}], bridges: [{id, name, shared_by, total_joint}], cooc: [{a, b, n}], degraded?}. degraded = the lookup failed: an empty graph is missing data to retry, not "no collaborators". More: omniseek docs/TOOLS.md, omniseek_coauthors.
Inputs: authors (required), hints, papers, source.
omniseek_curator_actA
Use WHEN acting on the source-curation protocol: WRITE a source-lifecycle action (submit / probe / decide / admit / retire ...); every safety gate lives in the impl, unchanged. Pick
the action with verb; the other arguments are that verb's:
submit (name, urls, mode, domain, family; optional kind, regions, rationale, draft): add a candidate; mode is STRUCTURE / UNWALL / TRANSCRIBE / RECALL / MONITOR.
probe (candidate_id): gather the evidence; returns the packet.
wall_probe (candidate_id): re-probe by rendering it in the isolated browser (a parked_p2 candidate, or one you judge a client-rendered shell).
decide (candidate_id, decision, reasons; baseline_ref to admit): admit / watch / reject. Refused: an admit with a hard red line, incomplete evidence, an empty baseline_ref or no packet.
apply_live (candidate_id): reversible live admit, rss-safe families only (others use stage_commit).
rollback_live (name, family): undo apply_live.
stage_commit (candidate_id): prepare the git commit text for a non-auto family; the operator commits.
retire_live (name; confirm): needs a PRUNE verdict; confirm=False previews, True writes a reversible override.
rollback_retire (name): undo retire_live.
source_verdict (name, verdict, rationale; prune_class, coverage_impact): KEEP / WATCH / PRUNE an existing source. Refused: a PRUNE its safety flags forbid.
Tool name: omniseek_curator_act (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_curator_act".
An unknown verb returns an error listing the valid values. Protocol: the /curator skill. More: omniseek docs/TOOLS.md, omniseek_curator_act.
Inputs: verb (required), kind, mode, name, +14 more.
omniseek_curator_viewARead-only
Use WHEN running the source-curation protocol (judge the admission queue or a source audit); READ curator state: queue | packet | audit. Never mutates. Pick a view with what:
what: "queue" (the admission backlog) | "packet" (the last evidence packet of a candidate, with the submitter's draft) | "audit" (per-source neutral audit facts and safety flags, no verdict).
candidate_id: the candidate, for what="packet".
state: filter for what="queue": new / probed / awaiting_verdict / admitted / watching / rejected / owner_review / redline_blocked / parked_p2 / error.
Tool name: omniseek_curator_view (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_curator_view".
An unknown what returns an error listing the valid values; a candidate with no packet yet returns {error} or {state}. Protocol: the /curator skill. More: omniseek docs/TOOLS.md, omniseek_curator_view.
Inputs: what (required), state, candidate_id.
omniseek_field_skeletonARead-only
Map a research field's shape: use WHEN you need its citation neighborhood (foundational core by citations vs frontier by date) to cluster yourself, from a topic or seed papers.
Tool name: omniseek_field_skeleton (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_field_skeleton".
Args: query: topic words; the top-relevance papers become the seeds. seeds: OpenAlex work ids you chose as anchors (preferred once you know the field). n_seeds: how many seeds to take from query. citers_per_seed: citing papers fetched per seed. source: "openalex" (default) or "s2" (Semantic Scholar: better arXiv coverage; nodes add influential, intent, contexts). max_nodes: node cap. fresh: bypass the cache. deadline_s: wall-clock cap (default about 25 s).
Returns {seeds, n_nodes, n_edges, edges: [[citer_id, cited_id]], nodes: [{id, title, year, date, cited_by, in_degree, concept, first_author, doi, url, is_seed, seed_ref_freq, seed_cite_freq}], _meta: {seed_titles, seed_note, degraded, deadline_hit, partial}}. in_degree = citations from inside the field; seed_ref_freq / seed_cite_freq on a non-seed node flag what your seeds miss. _meta.deadline_hit true = a partial map: retry or raise deadline_s. More: omniseek docs/TOOLS.md, omniseek_field_skeleton.
Inputs: fresh, query, seeds, source, +4 more.
omniseek_gatherARead-only
Run N independent read-only eye tools IN PARALLEL, returning results in one response.
Tool name: omniseek_gather (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_gather".
Args: calls: [{"tool": "omniseek_search", "args": {...}}, ...]: at most 10, read-only tools only; a call that needs another's result goes in a later gather. wait_s: return when all finish or after this many seconds; unfinished calls report status "warming" and keep running (collect later with staleness="cache_only", or repeat the same omniseek_transcribe).
Calls through one rate-gated upstream (several arXiv reads, say) run one after another at its pace; more than 2 search-index venues in one gather adds a warning. Returns {results: [{index, tool, status, result | error | hint}], elapsed_s, completed, warming, failed, total}; a wrong argument's hint names the tool's real parameters. More: omniseek docs/TOOLS.md, omniseek_gather.
Inputs: calls (required), wait_s.
omniseek_graphARead-only
Use WHEN you want HOW two entities connect, or what OmniSeek already knows AROUND a paper / author / entity; read-only, budgeted projections of its accumulated evidence graph with typed edges (ONE graph).
Tool name: omniseek_graph (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_graph".
Pass view and args (that view's own parameters); a call with no view lists the views and their parameters. Views return node ids, labels and edge tuples, never content. find {label_query, kind?}: the entry point, a name to candidate node ids. stats {}: counts by kind, type and tier. neighborhood {anchor, depth? (<= 2), types?, policy?, max_nodes?}: the subgraph around a node. between {a, b, types?, policy?, max_nodes?}: up to 8 shortest connection paths. voices {doc_ids (<= 64), policy?}: collapse documents to distinct upstream voices. since {anchor, date, types?, max_nodes?}: stored edges added after a date. similar {anchor (a doc), k?}: vector-nearest document candidates, proposals only. policy: conservative (exact-id identity only, default) | working (plus your omniseek_ruling rulings) | exploratory (plus fuzzy candidates). A cut result carries capped: true; a failure returns an error dict. Entity kinds fill in as calls happen, so early emptiness is normal. Record identity with omniseek_ruling, relations with omniseek_statement. More: omniseek docs/TOOLS.md, omniseek_graph.
Inputs: args, view.
omniseek_institution_cohortARead-only
Use WHEN you need the people-ROSTER of a lab / department / university (who actively publishes there, optionally scoped to a field): the "who's at this lab" question, orthogonal to co-authorship ("same lab, never co-authored" is still a tie, and co-authorship cannot see it).
Tool name: omniseek_institution_cohort (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_institution_cohort".
Args: institution: the lab, department or university. concept: the field to scope to ("machine learning", ...); without it you get the most prolific people across all fields. year_from: count works from this year on (biases to the current cohort). limit: how many people.
Returns {institution: {id, name}, filters, n, people: [{id, name, works_at_institution_in_field}], note}: a starting roster from OpenAlex, not a verified member list. More: omniseek docs/TOOLS.md, omniseek_institution_cohort.
Inputs: institution (required), limit, concept, year_from.
omniseek_paper_enrichARead-only
Use WHEN you need ONE paper's open-access full-text PDF, retraction / integrity status, or citation count: signals omniseek_search / field_skeleton do NOT give cleanly.
Tool name: omniseek_paper_enrich (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_paper_enrich".
Args: ids: DOIs and/or arXiv ids (not OpenAlex W-ids).
Returns {results: [{id, kind, doi, is_oa, oa_status, oa_by, pdf_url, oa_url, citation_count, integrity: {retracted, notices}}]}; an unrecognized id comes back as {id, error}. is_oa false = no free copy known to Unpaywall (still try omniseek_read on the DOI URL); retracted null = not checked; citation_count null = backend unreachable. Read pdf_url with omniseek_read for the full text. More: omniseek docs/TOOLS.md, omniseek_paper_enrich.
Inputs: ids (required).
omniseek_paper_recommendARead-only
Use WHEN you have a paper and want more like it: semantically-similar papers (SPECTER embeddings) that keyword search and the citation graph miss, including very recent work.
Tool name: omniseek_paper_recommend (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_paper_recommend".
Args: ids: seed paper ids: arXiv ids, DOIs or S2 ids (for an openalex hit, metadata.paper_id or metadata.doi, not the W-id). limit: how many papers.
Returns {seeds, n, papers: [{id, title, year, date, cited_by, first_author, doi, url}]} in S2 relevance order. More: omniseek docs/TOOLS.md, omniseek_paper_recommend.
Inputs: ids (required), limit.
omniseek_readARead-only
Read text from any URL OR document FILE: OmniSeek's single "read this deep" verb. AUTO-ROUTES.
Tool name: omniseek_read (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_read".
A local path, or a name ending in .pdf/.pptx/.docx/.xlsx/.txt/.md/.csv (and a local .html/.htm), goes to the document reader; anything else to the URL reader. arXiv /abs/ gives metadata, /pdf/ or /html/ the whole paper. A file from another machine goes into OmniSeek host's inbox ~/.omniseek/inbox first (or pass its URL).
Args: target: the URL or document path. start_char: where the returned window starts. max_chars: window size (default 24000); when truncated, re-call with start_char + returned_chars. level: "full" (default) | "brief" (title, author, abstract, size) | "head" (section list with char ranges) | "section". section: a section title (case-insensitive prefix) or index from level="head"; implies level="section". grep: a term, comma-separated terms or a list: return only the sentences that contain one, with one sentence of context. grep_word: grep matches whole words only. grep_case: grep matches case. export_media: (documents) export the embedded images. ocr: (documents) OCR every embedded image and append its text.
Returns (URL) {url, matched, total_chars, returned_chars, start_char, truncated, cached, fetched_at?, text_quality?, document, reason?}: matched false carries reason walled (drill the source by name with omniseek_search) | empty | blocked. Returns (document) {source, total_chars, returned_chars, start_char, truncated, cached, format, title, outline, media_total, media, media_dir, ocr_images?, text_quality?, text} or {source, error, inbox_files?}. text_quality (a PDF whose text layer reads as garbled): {level "low", readings, thresholds, hint}; read those pages with omniseek_view render_pages, and a term not found in that text may still be in the paper. With grep the body is replaced by grep: {terms, hits, occurrences, matches: [{start, end, text, terms}], returned_matches, truncated, next_start_char?}. An unknown section returns error and available_sections. To see figures use omniseek_view. More: omniseek docs/TOOLS.md, omniseek_read.
Inputs: target (required), ocr, grep, level, +6 more.
omniseek_resolve_identityARead-only
Resolve a PERSON's name to candidate author ids: the shared front door for EVERY relationship tool. It never picks: it returns ranked candidates for you to disambiguate.
Tool name: omniseek_resolve_identity (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_resolve_identity".
Args: name: the person's name. hint: an institution or field; only re-orders candidates. source: "auto" (OpenAlex, plus Semantic Scholar when the top hit is sparse), "openalex" or "s2". paper: arXiv id / DOI / title of a known paper by this person; pins a common-name junior.
Returns {query, source, candidates: [{id, source, name, works_count, cited_by, institution, via_paper?}], ambiguous, note, likely_same_person?: [{source, ids, name, merge_token, note}], degraded?}. ambiguous true = confirm with a hint, a paper or a known coauthor. merge_token ("A1+A2") goes straight into omniseek_coauthors. degraded = the OpenAlex lookup failed, so an empty result is missing data: retry, or pass source="s2" or paper=. More: omniseek docs/TOOLS.md, omniseek_resolve_identity.
Inputs: name (required), hint, paper, source.
omniseek_rulingA
Use WHEN two graph nodes ARE (or are NOT) the same person / entity and you want views to collapse them: record / list / retract same_as | not_same_as rulings (the one judgment channel the graph's working policy applies).
Tool name: omniseek_ruling (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_ruling".
create (src, dst, verdict "same" | "not_same"; optional note): returns {created, ruling, replaced}; the unordered pair is the key, so re-creating replaces it. list: {rulings: [{src, dst, verdict, note, ruled_at}], count}. delete (src, dst): {deleted}. omniseek_graph applies rulings under its working and exploratory policies; not_same beats same. A bad verdict, empty or identical endpoints, or an unknown action return {error}. More: omniseek docs/TOOLS.md, omniseek_ruling.
Inputs: action (required), dst, src, note, +1 more.
omniseek_searchARead-only
Search the curated sources. The default for "best/latest on X". ONE verb, three shapes.
Tool name: omniseek_search (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_search".
Shapes: default = one ranked, deduped, cross-lingual list; raw=True with exactly one source = an unbounded drill of that source (the way to reach walled or slow sources, which the broad sweep drops at its deadline); raw=True otherwise = uncollapsed per-source buckets.
Args: query: search words; empty browses by recency. sources: source names (None = every non-explicit_only source); take names from omniseek_sources. limit: max documents (per source in buckets). semantic: None = lexical plus vector recall; False = exact tokens only; True = lean on vectors. raw: see shapes. full: whole content per document. wait_s: patience budget in seconds (None = default deadline). staleness: "fresh" (bypass cache) | "cached_ok" (default) | "cache_only" (read what already warmed, send nothing). debug: keep the ranking telemetry in each document's metadata. filters: dict ({"year": {"gte": 2020}}) or expressions ("k=v", "k=a,b", "k>=n", "k=a..b", "k~text"); fields per source in omniseek_sources. as_of: "YYYY-MM-DD" or an ISO time with zone: only items that existed then.
Fire then collect: call with a small wait_s, then again ~20 s later with staleness="cache_only", the SAME query and the SAME limit (limit is in the cache key): faster and more complete.
Returns {query, count, documents, _meta, routing_hint?} (ranked); {source, query, count, documents, _meta?} (one-source drill: _meta.diagnostic when empty or errored, _meta.held when the eye's own gate sent nothing); {query, results: {source: [...]}, total_count, _meta} (buckets). _meta names timed_out / empty / errored / held sources (held = not tried, says nothing about the topic) and excluded_relevant walled sources worth naming. A bad filter returns error, filter_problems and available_filters and sends nothing; a bad as_of returns error. When _meta.web_search_backend shows cooling, wait cooling_s before retrying. More: omniseek docs/TOOLS.md, omniseek_search.
Inputs: query (required), raw, full, as_of, +7 more.
omniseek_sensorA
Use WHEN you want to MONITOR a query over time and be told only what's NEW: standing queries with novelty detection. ONE verb; action picks what to do.
Tool name: omniseek_sensor (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_sensor".
create (query; optional sources, schedule hourly | daily | weekly, notify, notify_if, notify_if_match "any" | "all", detect_absence, notify_source): register a sensor; returns it with its id (and schedule_note when the deployment needs a SERVICES.tsv row to run it). update (sensor_id, notify_source): change its push source ("" = default); returns {updated, sensor}. list: all sensors with last-run stats. delete (sensor_id): returns {deleted}. run (sensor_id): run it now; returns new_count and new_titles. notify=True pushes new results of scheduled runs (notify_if narrows that to keyword matches); detect_absence also alerts when a tracked stable-source item disappears. An unknown action or a missing required argument returns {error}. More: omniseek docs/TOOLS.md, omniseek_sensor.
Inputs: action (required), query, notify, sources, +6 more.
omniseek_sourcesARead-only
List all sources: call this to ROUTE before searching.
Tool name: omniseek_sources (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_sources".
A bare call returns the routing vocabulary and the bare inventory, not every source's facets.
Args: domain: only sources whose domains facet has it ("jobs", "papers", ...), with descriptions. query: sources ranked by token overlap with these words (English and Chinese), with descriptions. region: only sources whose regions facet has it ("sg", "ca", "cn", ...). verbose: every source with its full facets and description. field: with a query/domain that picks one source, also return that filter field's known values. check_health: probe every source live (slow) and add a system block.
Returns {count, backend_count, backend_breakdown} plus either the bare orient {source_names, available_domains, available_regions, capabilities, note} or sources: [{name, backend, description, needs_credentials, explicit_only, explicit_only_reason?, param_hint?, stability, access_tier, health, kind?, domains?, regions?, modes?, filters?, coverage?, license?, samples?, as_of_field?}]. backend_count counts distinct upstreams (count over-states coverage). health: ok | down | unmeasured | unverified | unknown | sealed; with check_health, healthy null means not verified, neither working nor failed. A near-miss domain or region returns did_you_mean. More: omniseek docs/TOOLS.md, omniseek_sources.
Inputs: field, query, domain, region, +2 more.
omniseek_statementA
Use WHEN you've concluded a DIRECTED, decision-relevant relation OmniSeek does NOT already store mechanically (X acquired_by Y, paper P refutes claim Q, path R requires gate S) and want the graph to carry it forward: record / list / retract typed relation statements (the general sibling of omniseek_ruling; identity types belong to omniseek_ruling).
Tool name: omniseek_statement (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_statement".
Record one only when all three hold: OmniSeek does not already store it (cites, authored, affiliated and the like are mechanical), it would change a future decision, and a point-in-time relation carries its as-of date in the note. Prose conclusions belong in notes, not here. create (src, dst, type, note; optional doc): type is free vocabulary, slugged to [a-z0-9_], at most 40 characters; note (your reasoning) is required; doc is a provenance node id. Returns {created, statement, replaced, similar_anchors?}: reuse a hand-minted id from similar_anchors instead of minting a near-duplicate. (src, dst, type) is the key; re-creating replaces it. list (optional about, type): {statements, count}, at most 200 (capped flag). delete (src, dst, type): {deleted}. Shown by omniseek_graph under working and exploratory, and on later omniseek_search hits of an endpoint. Refused with {error}: type same_as / not_same_as (use omniseek_ruling), a bad type, an empty endpoint or note, an unknown action. More: omniseek docs/TOOLS.md, omniseek_statement.
Inputs: action (required), doc, dst, src, +3 more.
omniseek_transcribeARead-only
Transcribe the SPOKEN content of a video / podcast / audio URL via local SenseVoice ASR (bilibili, xiaoyuzhou, direct audio). YouTube captions come from omniseek_read instead.
Tool name: omniseek_transcribe (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_transcribe".
Slow on the first call for a long item, then cached. For a long episode take the chapter from the shownotes and transcribe only that slice.
Args: url: the video, podcast or audio URL. language: "" auto-detects; "zh" / "en" skip detection. start: slice start, seconds or MM:SS / HH:MM:SS. duration: slice length, same forms. segments: also return [{start, end, text}] per segment (an extra pass). diarize: who said what: segments carry speaker labels (Chinese-focused, heavier; supersedes segments). speakers: known speaker count for diarize (0 = auto; pass it when you know it).
Returns {url, transcript, chars, audio_seconds, asr_seconds, source, title, cached, start_seconds?, duration_seconds?, segments?, speakers?}, or {url, error, transcript: ""} when no audio resolved. Repeating the same call while it runs attaches to that run; still running gives {status: "running", already_running: true, running_seconds, phase, note}: call again later. Over the machine's transcription limit it is queued and runs without you: {status: "queued", queue_position, note}; call again later with the same arguments. More: omniseek docs/TOOLS.md, omniseek_transcribe.
Inputs: url (required), start, diarize, duration, +3 more.
omniseek_viewARead-only
SEE with your own vision, IN-BAND: document figures, loose image URLs, or video frames.
Tool name: omniseek_view (your MCP config adds its own prefix; never guess one). Not in your tool list? ToolSearch "omniseek_view".
kind="auto" sends a document path or extension to document figures, a video URL (youtube, bilibili, douyin, .mp4/.webm/.mov) to frames, anything else to image URLs.
Args: target: document path or URL, video URL, or image URLs separated by commas, spaces or newlines. kind: "auto" | "document" | "images" | "video". sections: (document) slide or page numbers, comma-separated; "" = all. names: (document) exact image names from omniseek_read's outline media[].name. render_pages: (PDF) whole pages to render, for vector figures and tables with no embedded image. max_images: full-resolution cap; a wider selection comes back as a contact sheet. contact_sheet: (document) always return one tiled contact sheet. start: (video) slice start, e.g. "8:30" or "90". duration: (video) slice length; default the whole video, capped at 30 minutes. n: (video) frames to sample (default 12, max 24).
Returns image blocks: a contact sheet with a legend, or a manifest plus one block per image; or an error or "nothing to show" note. Usual flow: omniseek_read for the outline, this with no selection for a contact sheet, then the figures that matter by sections or names. More: omniseek docs/TOOLS.md, omniseek_view.
Inputs: target (required), n, kind, names, +6 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 18 tools
Most tools have clearly distinct targets: search/read/view/transcribe for content access, sensor for monitoring, curator_view/curator_act for read/write curation, and graph/ruling/statement for knowledge-graph operations. Some boundaries require careful reading — search vs paper_recommend vs field_skeleton, read vs view, and graph vs field_skeleton — but the descriptions generally clarify when to use each.
All names use the same omniseek_ prefix and lowercase snake_case, which is predictable and readable. However, the set mixes noun names (sources, sensor, graph, statement), verb names (search, read, transcribe, gather), and inconsistent compound patterns (paper_recommend/enrich vs resolve_identity, curator_view/act).
Eighteen tools is heavy for a single server and sits in the borderline 16–25 range, especially with many specialized sub-areas. The ambitious research/discovery/curation scope justifies the breadth, but the surface still feels large enough to burden tool selection.
The surface broadly covers discovery, reading, media transcription/vision, paper enrichment/recommendation, people and institution lookup, monitoring, graph reads/writes, and curation lifecycle. Minor workflow gaps may remain, but most core research tasks have a follow-through path rather than an obvious dead end.
Maintenance
Related MCP Connectors
An MCP memory server. One memory your agents share — across models, devices and apps.
Search engine for AI agents to find MCP servers, A2A agents, and skills on their own.
One MCP for any agent: your connected apps, shared memory, web search, social data and media tools.
Web search API and MCP server for developers building AI agents.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceMCP server giving AI agents real-time web search, page scraping, company intelligence, email discovery, local lead generation, and a persistent knowledge graph. Pay only for what you use, no subscriptions.MIT

Augentofficial
AlicenseBqualityDmaintenanceMCP server that turns any audio or video source into structured, searchable intelligence for agents, enabling download, transcription, semantic search, speaker identification, and more.226MIT- AlicenseNot gradedqualityBmaintenanceA private, self-hosted MCP server that wraps a retrieval pipeline over your own data, enabling trusted AI agents to access and manage your personal memory through standard MCP tools.Apache 2.0
- AlicenseNot gradedqualityAmaintenanceMCP server that enables AI agents to watch and analyze videos from 1800+ sources, with persistent indexing, OCR, transcription, and a self-verification loop for debugging.MIT