Context logo
Agents Orchestration Frameworks · Workflow Automation

Context

Context is an enterprise platform for building, running and improving AI agents on real business workflows. Its Workspace, Engine, Unify and Evals modules run on your own infrastructure—managed, VPC, on-premises or air-gapped—with 800+ permissioned connectors.

Active GDPR compliant Usage Based No public API 13+ Verified by Guidaio
Overview

What is Context?

Context is an execution platform for enterprise AI agents, published by Explore Interfaces Inc. in San Francisco. Where a chatbot answers a prompt, Context runs a defined sequence of steps across the systems a team already uses, with a written specification, permissions, review points and a definition of done. The product is made of four modules that can be deployed separately or together. Workspace is where people and agents work on the same files: runbooks written in plain language, one sandboxed computer per agent, applets generated on demand, a replayable trace and a per-task choice of model. Engine is the permissioned runtime: identity inherited from the customer's identity provider, a sandbox per session, 800+ connectors, authorisation before each action, severity gates and an append-only audit log, with durable runs that can be resumed. Unify keeps institutional context—documents, procedures, accepted examples and corrections—in a file system agents can browse. Evals scores each run against rubrics and golden sets and blocks regressions before a change reaches production. Deployment is the difference the company insists on: managed, customer VPC on AWS, Azure or GCP, on-premises or air-gapped, and it is the whole platform that ships, not only inference. Context is model-agnostic—Claude, GPT, Gemini, Kimi, DeepSeek, GLM or open weights—and can route each step to the cheapest model that passes the rubric. Context Inference adds managed inference inside the VPC, with private endpoints and distilled models the customer owns. Beyond the web workspace, Context is reachable from Slack, Microsoft Teams, the context-code CLI and a desktop client in private preview, alongside Agent Sandboxes, Agent Identity and Context Wiki. A marketplace lists 78 prebuilt agents and connectors across nine verticals, from semiconductors and financial services to insurance BPO and legal. The homepage claims a 40x cut in processing time and a 28x drop in cost per case; both come from an internal F100 benchmark, averaged over three runs, not from an independent third party. Usage is metered in Context Compute Units, decoupled from the model provider.

What it does

  • Write a runbook in plain language and have an agent execute it from end to end
  • Connect enterprise systems—warehouses, documents, CRM, ticketing—through 800+ permissioned connectors
  • Give agents the identity and access rights of the user, inherited from Okta or Azure AD
  • Produce editable deliverables: documents, spreadsheets, slide decks, kanban boards, code and meeting notes
  • Score every run against rubrics and golden sets, and block regressions before release
  • Route each step to the cheapest model that still passes the rubric
  • Run the whole platform in your VPC, on-premises or fully air-gapped, and schedule recurring runs
Audience

When to use Context / When not to

A quick filter to help you decide if Context is the right fit.

When to use Context

  • Operations teams running repeatable workflows that still depend on documents, several systems and expert judgment
  • Security-constrained organizations that need a VPC, on-premises or air-gapped deployment, public sector included
  • Financial services, consulting and legal teams handling due diligence, case processing, research and reporting
  • Engineering and industrial groups in semiconductors, telecom or manufacturing automating engineering operations
  • Founders and small teams looking for a single multi-model workspace instead of a stack of separate AI subscriptions

When not to use Context

  • Buyers who want a published price and a self-service sign-up: everything goes through a quote and a scoping call
  • Mobile-first users, since there is no iOS or Android app and Context Desktop is still a private preview
  • Developers looking for public API documentation or a developer portal, as neither is published
  • Procurement teams that require a published SOC 2 or ISO attestation before they will even start an evaluation
  • Teams wanting an instant, zero-setup assistant: connectors, permissions and an evaluation set have to be scoped first
Get started

How to use Context

A typical end-to-end flow, from setup to results.

  1. Start on the get-started page: describe the task and pick the deliverable type—nothing runs from the landing page
  2. Sign in and review the prepared task, changing whatever calls for human judgment
  3. Send it deliberately: the work only begins on an explicit Send
  4. For an enterprise evaluation, use the contact or guided demo form and bring a real workflow
  5. Let Context map the systems, permissions and quality controls, then agree on the deployment model
  6. In the Workspace, write the runbook in plain language, connect the sources, launch the run, replay the trace and edit the deliverable
  7. Connect each source once; agents then read, work and write back within the access rules you have set
  8. For the CLI, run the published curl install command, then context-code login and approve the displayed pairing code
  9. Add Context to Slack, mention the agent in a channel and approve steps in thread; results come back to the workspace
  10. Request private preview access for Context Desktop, then schedule recurring runs with cron expressions
Quick read

Pros & Cons

Pros

  • The whole platform deploys inside the customer's perimeter, not only the inference layer
  • Model-agnostic, open weights included, with per-step routing on cost against a quality rubric
  • Identity inherited from the identity provider, authorisation before each action and an append-only audit trail
  • 800+ permissioned connectors, with credentials never written into the runbook or the prompt
  • Evaluation is built in through rubrics and golden sets instead of being bolted on from a third party
  • Complete public legal documentation: terms, privacy policy, DPA, acceptable use policy, named subprocessor list and trust center
  • Plain-language runbooks a business owner can write and review, and an MIT-licensed open-source CLI base with published checksums

Cons

  • No public pricing and no pricing page: budgeting is impossible without going through sales
  • The trust center's Certifications page reads "Being populated": no SOC 2 or ISO attestation is published
  • Key security documents—pentest, CAIQ, security policies, insurance certificate—are available on request only, sometimes under NDA
  • No customer is named publicly; case studies are announced as coming
  • The 40x and 28x figures come from an internal benchmark, not from an independent third party
  • No public API documentation, no developer portal, no mobile app, and Context Desktop is still in private preview
  • Every subprocessor is US-based, no EU hosting option is announced for the managed mode and no Article 27 EU representative is designated
Pricing

Pricing & Plans

Context publishes no price. There is no pricing page on the site and none in its sitemap, so no starting amount and no currency can be stated here. The FAQ indicates that pricing depends on the deployment model, usage, support requirements and implementation scope, and directs prospects to a quote through the contact form. Usage is metered in Context Compute Units, presented as decoupled from the model provider, and CLI sessions are billed as workspace task usage on the team account. The founders page shows "Try for free" and "Free setup" wording, but neither a permanent free tier nor a time-limited free trial is documented anywhere on the site. The dollar amounts shown on the CLI page are third-party model provider rates, not the cost of Context.

No priced plan is published
  • the site shows no tier
  • no price grid and no per-plan feature table
Plan 3
  • A separate offer for founders and small teams is announced on the founders page
  • again without any price
Special offers — No promotion, discount, coupon or student, non-profit or jobseeker rate is published on the site · The closest thing is the founders page, which advertises "Free setup" assistance for founders and small teams, together with the offer to run a real workflow during a guided demonstration; neither is a priced discount, and neither documents a free trial
Prices and plans listed above may evolve. Always check the official pricing page before subscribing.
Trust & Privacy

Data, GDPR & hosting

A consolidated view of how Context handles your data.

GDPR overview

GDPR implementation is documented rather than merely claimed. A data processing addendum, last updated on 6 July 2026, is published without authentication. It covers EU GDPR 2016/679, UK GDPR, the Swiss Federal Act on Data Protection and CCPA/CPRA. The customer acts as controller, Context as processor. International transfers rely on the EU standard contractual clauses (Decision 2021/914, modules 2 and 3), the ICO's UK IDTA addendum and Swiss adaptations, the addendum's Annexes 1 and 2 serving as SCC Annexes I and II. Context commits to breach notification without undue delay, DPIA assistance, ten days' notice before adding a subprocessor and one audit per twelve-month period. Two limits: no EU representative is designated under Article 27, and the privacy policy has no detailed data subject rights section, only privacy@context.ai to update your data.

Who owns the data?

The terms state that, as between the two parties and to the extent permitted by applicable law, the customer retains all ownership rights in Inputs and Outputs. The customer grants Context a non-exclusive, royalty-free, worldwide licence to reproduce and use Customer Data solely as far as necessary to provide the service. Under the published data processing addendum the customer is the controller and Context the processor. In the VPC, on-premises and air-gapped modes, run traces stay inside the customer's own deployment and nothing is sent to Context. Distilled models trained on accepted work are presented as models the customer owns and serves itself.

Reuse rights

Outputs belong to the customer, who can reuse and commercialise them without asking Context for permission, subject to two contractual limits set by the acceptable use policy: Outputs may not be used to train a competing model, and the service may not be used for competitive analysis. On the vendor side, the FAQ states that one customer's traces, corrections or institutional context are never used to train models for other customers, and the trust center lists a "No cross-customer training" control. The terms still let Context use and internally modify Customer Data to deliver the service, and Usage Data telemetry is collected. Retention and training settings of the underlying model providers depend on the endpoints the customer selects, documented and checked at deployment. Outputs accepted by the team can become training data for distilled models that the customer owns. Website data collected on context.ai (name, email, phone, company, IP address, resume) is a separate matter: it is used for the service, marketing, analytics and security, with third-party advertising cookies and pixels and Google's invisible reCAPTCHA.

Data retention & training

Retention summary
Retention comes from the DPA and the terms rather than from a published schedule. On expiry or termination, and at the customer's election made within thirty days, Context deletes or returns all customer personal data in its possession and deletes the existing copies. Two exceptions apply: data the law requires to be kept, and data sitting in routine backups, which is then removed on Context's standard deletion schedules and protected in the meantime. The terms add that the service may impose storage duration and volume limits, beyond which Customer Data can be deleted automatically, and that Context may delete all Customer Data at the end of the contract. Export and retention requirements are configured per deployment, and run traces stay inside the customer's own deployment. No figure is published: neither the addendum nor the privacy policy says how long data is actually kept.
Trains on customer data
No
Subprocessors disclosed
Yes
DPA available
Yes
GDPR contact

Hosting summary

Where the data lives depends on the deployment mode, and four are offered: Managed, operated by Context; Your VPC, inside the customer's own AWS, Azure or GCP account, under its controls and its keys; On-premises, in the customer's data centre; and Air-gapped, fully disconnected. In the VPC, on-premises and air-gapped modes, the control plane, execution and run history all stay within the customer's perimeter. For the managed mode, the published subprocessor list is entirely United States-based: AWS for hosting, Microsoft Azure for infrastructure and inference, Vercel for the website and edge, with Datadog and Grafana Labs added for monitoring on 9 June 2026. The model providers listed are Anthropic PBC and OpenAI LLC; when a customer connects its own provider account, that provider is no longer a Context subprocessor. Data is encrypted with TLS in transit and industry-standard encryption at rest. The privacy policy states that the company is located in the United States and that data may be stored in any country where Context or its providers operate.

Hosting countries
🇺🇸 United States
Availability

Where Context works

Country-level availability.

Not available in

No country or region is excluded on the site. The terms carry only a generic United States export control clause, forbidding export of the service to embargoed or otherwise restricted jurisdictions, without naming a single country.
Watch-outs

Things to keep in mind

Risks and trade-offs to weigh before adopting Context.

  • No published price: the total cost depends on the deployment mode, on Context Compute Unit consumption and on implementation scope, which makes budget drift easy to miss
  • The trust center's Certifications page was still marked "Being populated" on 4 September 2026, and the pentest, CAIQ, security policy and insurance documents require a request and sometimes an NDA: you may be trusting statements you cannot yet verify
  • The 40x and 28x gains come from an internal F100 benchmark that no third party has audited; treat them as a vendor claim, not as a forecast for your own workflows
  • In managed mode every subprocessor sits in the United States, model providers included, so personal data leaves the EU under standard contractual clauses
  • The terms say the service is not designed to store sensitive data and the customer undertakes not to submit any, while the acceptable use policy forbids using Outputs to train a competing model or the service for competitive analysis
  • Agents acting across production systems amplify a badly written runbook: if approval points become a formality, a team can rubber-stamp work it no longer understands and quietly lose the expertise the agent replaced
  • No support email address is published and Context Desktop remains a private preview, so help goes through a form or the sales team; note too that the 2016 Wayback capture of the domain says nothing about the publisher's actual age
Setup

Setup & Integrations

Technical difficulty

Two very different paths. Self-service is easy: creating a workspace takes about a minute, the get-started page prepares a task without an account, and the CLI needs one install command plus a pairing code, with an automatic SHA-256 check. The enterprise path is a project: workflow, systems, permissions, review points and definition of done have to be scoped first, each connector authenticates through OAuth or a scoped service account, and an identity provider such as Okta or Azure AD is required. VPC, on-premises and air-gapped modes are infrastructure work with documented data-flow boundaries.

Deployment

Web appDesktop appSlack app

Integrations

Snowflake Databricks BigQuery Postgres Looker Box Notion Confluence Google Drive Salesforce HubSpot Zendesk ServiceNow Jira GitHub Slack Microsoft Teams Asana Linear Dropbox Okta Azure AD FactSet
Company

Behind Context

Company name
Explore Interfaces Inc.
Founded
09/04/2016
Country of origin
🇺🇸 United States
Headquarters
55 2nd St., Suite 1925, San Francisco, CA 94105
UBO
Joseph Semrai
UBO country
🇺🇸 United States
Domain registrar country
🇺🇸 United States
Legal contact

Fundraising

context.ai publishes nothing about its funding: no investor page, no press release and no mention in the press section. Everything below comes from off-site press coverage only and is not confirmed by the publisher
Off-site (TechCrunch, 28 May 2025): a seed round of 11 million USD at a 70 million USD valuation, led by Lux Capital with Qualcomm Ventures and General Catalyst
Off-site: total funding estimated at around 15.75 million USD
Off-site: the company is reported to have been founded by Joseph Semrai, a Thiel Fellow, who started working on Context in 2024

Social

Official links

Resources

All the official URLs gathered for verification and reference.

Compare

Alternatives

Tools that compete with or complement Context.

C ChatGPT EnterpriseC ClaudeC Claude CodeM Microsoft 365 CopilotG GleanZ Zapier AgentsC CodexC CoworkH HarveyH HebbiaR Rogo
FAQ

Frequently asked questions

What exactly is Context?
A platform for building and running AI agents on business workflows. It brings together four modules—Workspace, Engine, Unify and Evals—that can be deployed separately or together.
How is it different from a chatbot or a copilot?
Instead of answering a prompt, Context executes a defined sequence of steps across your systems, driven by a runbook that sets permissions, review points and a clear definition of done.
What is a runbook?
The operating specification of an agent. It records the steps to follow, the systems the agent may use, expected inputs and outputs, approval points and evaluation criteria.
Where does Context run, and how long does a deployment take?
Managed by Context, in your own VPC on AWS, Azure or GCP, on-premises, or in a disconnected air-gapped environment; the whole platform is deployed, not only inference. Timelines depend on your deployment requirements, connector access and the risk of the workflow, a first workflow covering the runbook, the systems needed, a small evaluation set and reviewer sign-off.
Do the agents hold our credentials?
No. Credentials are handed to the connector at run time and are never written into the runbook or into the prompt.
Will our data be used to train models?
Not for other customers: Context states it does not use one customer's traces, corrections or institutional context to train models for others. The retention and training settings of the model providers depend on the endpoints you choose.
Which models can be used?
Commercial or open-weight models, depending on the deployment. A runbook can pin a single model or route each step according to cost, latency and evaluation performance.
How much does it cost?
No price is published. Pricing depends on the deployment model, usage, support requirements and implementation scope, and is quoted on request through the contact form.
Is there a mobile app?
No. Context is used from the web workspace, Slack, Microsoft Teams and the context-code CLI, plus a desktop client for macOS and Windows that is still in private preview. There is no iOS or Android application.
Is a data processing addendum available?
Yes. A DPA is published on the site, with standard contractual clauses, the UK IDTA addendum and a named subprocessor list. The privacy policy also states that the services are not intended for anyone under 13.
Conclusion

Should you pick Context?

Context is an enterprise tool, not a consumer assistant. Everything about it—runbooks, permissions, review points, evaluation sets—is built around workflows a company already runs and would rather not hand to a generic chatbot. Its strongest arguments are architectural. The whole platform, not just the inference layer, deploys inside the customer's perimeter: managed, in your VPC, on-premises or fully air-gapped. Agents inherit each user's identity from Okta or Azure AD, ask for authorisation before every action and leave an append-only audit trail. Evaluation is part of the product rather than an afterthought, which matters as soon as an agent touches production systems. Model choice stays open, open weights included, and each step can be routed to the cheapest model that still passes the rubric. The reservations are just as concrete. There is no public price and no pricing page, so budgeting means talking to sales. The trust center's Certifications page is still empty, so no SOC 2 or ISO attestation can be checked today, and the security documents sit behind a request and sometimes an NDA. No customer is named. The 40x and 28x performance claims come from an internal benchmark. Every subprocessor is US-based, no EU representative is designated under Article 27, and no EU hosting option is advertised for the managed mode. Explore Interfaces Inc. is a young San Francisco company, yet its legal paperwork is already unusually complete for its stage: terms, privacy policy, a published DPA with standard contractual clauses, an acceptable use policy and a named subprocessor list. The sensible entry point is a guided demonstration built on a real workflow, or a first deliverable prepared through the get-started page. Ask for the compliance evidence and a written quote before you commit.