Arcness logo

Enterprise AI gateway

Connect your knowledge. Control your AI.

One gateway for every AI in your company. Model calls, company knowledge and coding tools all run under the same permissions, guardrails and records.

38connectors21model providers12guardrail enforcement points
Contact sales

Hover over or click the logo

ChatsNew chat
Security audit summary
Q3 sales report draft
New hire onboarding
Supplier contract comparison
Agents
Contract review
Security Q&A
Company policy assistant
Security audit summaryIn-house model
Summarize this year's security audit findings and their deadlines
Searched 14 documents in SharePoint
Searched 3 documents in Google Drive
Checked access permissions, 2 excluded
Answer complete

This year's audit raised 3 findings.

1. Delayed deprovisioning of departed employees: automatic removal by September 301

2. External sharing links with no expiry: 30-day default expiry by October 152

3. No backup recovery drill: one drill in November1

Ask a follow-up question
Source 1Source 2
2026 Security Audit Results.pdf
SharePoint / Security / Audit · p. 12
Accounts of departing employees must be revoked within one business day, but across 42 sampled cases revocation took 6 days on average.
Access: Security team, Audit committee
Model management · model types
ChatEmbeddingRerankerImageGuardNEWDiscriminative
System One API connections
JJevTypeSafeProvider API
JJevVercel AI GatewayGateway
JJevDigitalOceanCloud
OOpenJevCodivHosted
OOpenJevvLLMSelf-hosted
KKevOllama · vLLMSelf-hosted
New

Discriminative models, starting with Jev

Jev, released by TypeSafe in September 2026, is a new class of model: it generates no text and returns one of the defined options with a probability. Arcness adds discriminative models as a model type in model management to support it.

  • Decisions with probabilitiesYes/no, choice and graded-score answers come with a probability, and thresholds decide the action.
  • Applied where it mattersGuardrail AI evaluation, auto evaluation and the search reranker use it instead of a chat model.
  • One API, many connectionsConnect TypeSafe, Vercel AI Gateway, DigitalOcean and self-hosted OpenJev or Kev the same way.
Learn more
LLM gateway + α

An LLM gateway, plus
knowledge and control

A typical LLM gateway connects and logs model calls. Arcness also handles company knowledge, guardrails, coding tools and the work itself, through the same gateway.

Typical LLM gatewayArcness
Model connections✓✓ 21 providers
Token and cost control✓✓ Limits · budgets
MonitoringCall logsTraces · quality evaluation
Company knowledge+α–38 connectors · knowledge graph · document permissions
Guardrails+αPartial12 enforcement points · block · pseudonymize
Coding tools+α–Claude Code · Codex CLI · Cursor
Doing the work+α–Agents · harnesses (result files)
38Document sourcesConnectors for wikis, storage, messaging, ticketing, dev tools and more
21Model providersCloud, vendor API, gateway and self-hosted
13Personal data types detectedIncluding Korean formats such as resident and business registration numbers
12Guardrail enforcement pointsFrom chat input to files, APIs and the code gateway
Connectors

Scattered documents,
one knowledge base

Connectors collect and index documents from your company systems on a schedule. Agents answer from these documents and cite a source for every sentence.

NNotion
CConfluence
SSharePoint
GGoogle Drive
SSlack
JJira
GGitHub
BBox
Arcness
Answer with cited sources123

Connectors by category, with source permissions

Register a connector in the admin console and collection and indexing run automatically. Tokens are stored encrypted, and each document can follow the access permissions of its source system.

WikiStorageMessagingTicketingProjectsDevelopmentMailWebPublic dataResearch
38document sources10external databases5search engines
Real product screen
Connector catalog and registering the Notion connector
Grounded answers

From tens of thousands of chunks,
only the evidence you need

For every question, Arcness searches the document chunks within the user's permissions, reorders them with a reranker and keeps only the evidence the answer needs. Keep scrolling.

Answer with 3 cited sources
0Chunks searched
Knowledge bases · Databases

Ask documents and data
in natural language

Use policy documents as knowledge bases and business data through database connections. Agents draw on both to answer.

Real product screen

Knowledge bases and grounded answers

PDF and DOCX files are indexed on upload, and a search test lets you check the retrieved chunks first. Every sentence of an answer carries a source number.

A source number on every sentence
Real product screen

Natural-language database queries

Connect 10 databases including PostgreSQL, Oracle, Snowflake and BigQuery. Ask in natural language, get a chart, and verify with the executed SQL.

SQL execution and chart answers
Knowledge graph

Scattered knowledge,
one graph

Connect database schemas, operating rules and glossaries as nodes and edges. Agents follow the links to answer with the context between data and rules.

TableColumnConceptTermDocument

Select two nodes in turn to find the shortest path. Hover over a node to highlight its connections. Sample data from the example company ACME Logistics.

5node typesShortest paths · Cypher queriesApache AGE · Neo4j · Amazon Neptune
Real product screen
Finding a path from support tickets to return requests
Document-level access control

Answers limited to
what each user may see

Arcness can follow the permissions of the source system. Documents a user cannot access are left out of both the answer and the source list.

F
Finance team lead2,410 accessible documents
S
New sales associate312 accessible documents
Keep scrolling to switch the user asking
FSShow Q3 operating profit by business unit
Cloud
[Amount]
Solutions
[Amount]
Consulting
[Amount]
Training
[Amount]

No Q3 profit and loss data was found in the documents you can access.

Request access from Finance
Source: Q3 business results.xlsx (Finance shared folder)No sources (2 documents outside your access excluded)
Memory

What one person teaches
carries into the team’s answers

The agent gathers the working knowledge teammates share in chat into agent memory and uses it in everyone’s answers. A rule you save with “remember that” shapes only your own answers, and memories learned from a private source show only to the people it is shared with.

JJamie ParkOperations adminPrivate source shared
AAlex RiveraLogistics teamPrivate source shared
MMorgan LeeSales teamNot shared
Logistics Operations AssistantSample data
TopicPersonal memory · only youAgent memory · the team
Selected topic

    Sample graph · each node is one memory, the large nodes are topics

    Real product screen

    Agent memory the whole team shares

    The agent remembers a logging code a teammate shared and uses it in someone else’s answer, and memories learned from a private knowledge base show only to the people it is shared with.

    Memories show who taught them and how often they are used
    Real product screen

    Memory that learns from chat and shapes answers

    Save the weekly report format with “remember that”, find it in the topic graph, and the next chat’s answer follows it. Admins can clean up one user’s memories.

    A weekly report in the saved format
    Agent memory · shared by the teamLearn from chatsTopic node graphInherited permissionsRight-to-be-forgotten requests
    Harnesses

    Domain procedures,
    packaged as a harness

    A harness bundles instructions, skills, libraries and completion checks into one work environment. The model reads the attached files and runs code in an isolated container to produce result files, and the completion checks recompute those results.

    Start from a harness templateCopy a domain's instructions, skills and completion checks, then connect your own data to the resource slots.

    Cell image quantificationSample run
    8 microscopy images

    Compare cell counts and fluorescence intensity by treatment group against the SOP.

    Completion checks
    Result files

    A sample screen. Real steps and times depend on the model and the files.

    Real product screen

    Quantify cell images with a harness

    Keep the analysis SOP in the instructions and hand it eight microscopy images: nucleus outlines, a group comparison chart and a results table come back, and follow-up analysis continues in the same work folder.

    Nucleus outlines drawn over the originals
    Real product screen

    Drawing review from a harness template

    Copy the template, connect a material master and a design standard to its resource slots, then hand it a DXF drawing in chat and get back a marked-up drawing and a bill of materials.

    A marked-up drawing that passed the checks
    Harness templatesCompletion checksPick directly in chatLocal install for Claude Code · Codex
    Guardrails

    Sensitive data is protected
    before it reaches the model

    Rule-based detection, AI evaluation and prompt injection detection inspect every input and output. Depending on policy, data is blocked or pseudonymized, and every verdict is recorded in the guardrail logs.

    Chat inputTry editing it
    Action
    What the model receives

    Where checks run

    Every request flows through the detectors at each enforcement point. Hover over or tap a branch to see its verdict.

    Sample

    12 enforcement points

    Chat inputModel outputAttachmentsText in imagesTool resultsMemoryFlowsDB extractionFilesScheduled tasksAPICode gateway
    13personal data types6actionsConfidential document similarity detection
    Real product screen
    Contact details and card numbers reach the model as pseudonyms
    Security · Operations

    Ready for your
    security review

    Authentication, auditing, encryption and deployment: the items enterprise security teams check first. Move your cursor over the cards.

    Single sign-on and automatic provisioning

    Sign in with Microsoft Entra ID, Google Workspace, SAML 2.0, OpenID Connect or LDAP, with SCIM user sync and two-factor authentication.

    Sealed audit logs

    Records time, user, action, resource and IP. Each record carries the hash of the one before it, so integrity can be verified.

    Encryption key management

    Use the built-in key or integrate with Azure Key Vault, AWS KMS, Google Cloud KMS or Vault, with key rotation.

    Per-organization databases

    In installations that host several organizations, each organization gets its own database and dedicated role.

    On-premises deployment

    Docker Compose for a single VM and a Helm chart for Kubernetes. The same image installs in air-gapped networks and private clouds.

    Distribution channels

    Bring AI into company systems through a Microsoft Teams bot, embed widgets, an OpenAI-compatible API and MCP.

    LLM gateway · Code gateway

    Every model call,
    governed in one place

    Connect models from 21 providers across cloud, vendor APIs, gateways and self-hosted deployments through a single gateway. Requests from coding tools pass through the same policies and logs.

    Code gateway pipeline
    1
    AuthenticationDeveloper verified by issued key
    2
    Policy checkBlocked repositories, files and commands
    3
    GuardrailsSensitive data inspection
    4
    Allowed modelHandled by an administrator-approved model
    5
    Coding session logCall record and cost tracking
    Claude CodeCodex CLICursor

    Policy violations are logged as warnings by default, and administrators can switch them to blocking. Set which users and groups can use each model, along with daily token limits and pricing, to manage cost.

    21model providers4connection typesAccess, limits and pricing per model
    Real product screen
    Code gateway policies and allowed models
    Evaluation · Tracing

    Quality scores and
    how answers are generated

    An evaluation model scores answer quality, faithfulness and retrieval quality. Traces record search, model calls and tool runs by time and tokens, and can be exported with OpenTelemetry.

    TraceRun 7 · 445 ms · 2,599 tokens
    TreeTimelineTokens

    Question: How far in advance must I request annual leave? · Sample data

    LeaderboardIncludes blind evaluations
    #21,004Gemini 3 Flash
    #11,083Claude Opus 4.6
    #3962gpt-5.6-luna
    0%
    Answer quality91%
    Faithfulness88%
    Retrieval quality82%

    Average auto evaluation score · Sample data

    7trace span typesElo-based leaderboardArena with model names hiddenOpenTelemetry export

    Trace conversations as well as scheduled tasks, document indexing and coding tool calls on the same screen. Usage, estimated cost and audit logs sit in the same admin panel.

    Real product screen
    Inspecting the raw input and output of each span
    Discriminative modelsNEW

    Ready for the newest
    class of model

    Jev, released by TypeSafe in September 2026, is a discriminative model: it generates no text and returns one of the defined options with a probability. Arcness adds discriminative models as a model type, connects any model that exposes the same API, and applies them to guardrail AI evaluation, auto evaluation and the search reranker.

    Model management · model types
    ChatEmbeddingRerankerImageGuardNEWDiscriminative
    System One API connectionsThey share one API, so switching providers is a change of connection.
    JJevTypeSafeProvider API
    JJevVercel AI GatewayGateway
    JJevDigitalOceanCloud
    OOpenJevCodivHosted
    OOpenJevvLLMSelf-hosted
    KKevOllama · vLLMSelf-hosted
    3Decision typesYes/no · choice · graded score
    3FeaturesGuardrail AI evaluation · auto evaluation · search reranker
    6ConnectionsProvider API · gateway · cloud · self-hosted

    Latest jev-1.13.0 · version pinning

    Guardrail AI evaluationSample

    A yes/no probability is compared with two thresholds to decide the action.

    Yes/noDoes the content request or disclose a customer’s personal contact details?

    Give me the customer Minsu Kim’s mobile number.

    Block probability0.91
    Block

    Conventional AI evaluation returns only pass or block. Because a discriminative model returns a probability, requests between the two thresholds are not blocked but kept in the guardrail inspection record.

    Pass
    Log
    Block
    Log threshold 0.35
    Block threshold 0.70
    00.350.701

    Three decision types

    The input is read once and several decisions run at the same time. Every result comes with a probability, which each feature uses as a score from 0 to 1.

    Q. How many days ahead do I request annual leave?A. Request annual leave in the HR system at least 3 days before. [1]
    Graded scoreAnswer helpfulnessAuto evaluation
    Very lowLowMediumHighVery high
    3.4 / 4 → 0.85
    ChoiceMost relevant documentSearch reranker
    HR policy 5.20.74
    Leave FAQ0.18
    Benefits guide0.08
    Yes/noContradiction with the sourceAuto evaluation
    Yes0.04

    Uses assigned per model

    Connect external services and self-hosted models side by side, and assign the discriminative model each feature uses.

    Guardrail AI evaluationAuto evaluationSearch reranker
    JevTypeSafe · external service
    OpenJevSelf-hosted · GPU
    KevSelf-hosted · Ollama

    A newly connected discriminative model cannot be selected in any feature until it is assigned a use.

    Coming nextTool-call approvalTool selectionFlow router
    Yes/no · choice · graded scoreProbability thresholdsUses assigned per modelSystem One APISelf-hosted for closed networks
    Product videos

    See it in
    the real product

    Feature videos recorded from the real product, using sample data from the example company ACME Logistics and a sample lab.

    Easy document connections2:13

    Documents stay where they are; only the evidence comes to the answer. Permissions follow the source, and every action is on the audit record. Control of your data always stays with you.

    See it with
    your own documents

    In a consultation, we connect a few of your documents so you can see the answer quality for yourself. Leave your question below and our team will get back to you.

    Inquiry type

    We collect your name, company, email and message to answer your inquiry, keep them for one year after we reply, and then delete them. Without your consent we cannot receive the inquiry.

    Product videosReal product screens · Example company ACME Logistics · 11 videos