# Keld, Machine Version > Keld is the world's first **marketplace for AI inference**, connecting the enterprises that build apps and agents with the AI model providers that power them. Keld is the platform where a company first **understands** its AI work and then **optimizes** it. **Keld Atlas** is the understanding layer: it shows how a company uses AI and what it costs across **personal inference** (people working in AI tools such as coding assistants and chat apps) and **agentic inference** (the apps and agents the company builds on models), highlights inefficiencies, and lets leaders track usage and cost. Understanding is organized around **workstreams**, the lines of work a company declares (projects, marketing campaigns, customer efforts, GTM phases, engineering repositories); optimization works on **activity types**, the unit of work Keld routes and clears on the marketplace. Keld reads the work behind each prompt, classifies it into an activity type, and lets the company **re-route the activity types it chooses** (coding, text creation, extraction, summarization, Q&A) to Keld, where each is cleared on the best-value provider of the model that fits, within the buyer's bounds. Re-routing requires no code changes. **Quick links:** [Keld Atlas](https://keld.co/atlas) (free · for enterprises) · [Keld Marketplace](https://keld.co/marketplace) (private beta · for enterprises) · [Keld Trade](https://keld.co/trade) (private beta · for AI model providers) · [Documentation](https://keld.co/docs) · [Blog](https://keld.co/blog) · [Trust Center](https://keld.co/security) · [Join the Atlas private beta](https://keld.co/signup) **License:** © 2026 Keld. This brief and the content on keld.co may be quoted and indexed by AI assistants and search engines with attribution to Keld (https://keld.co). --- ## Overview - Keld connects the enterprises that build apps and agents with the AI model providers that power them, and runs each piece of work at the best price within the enterprise's bounds. - **Understand, then optimize.** Companies cannot manage what they cannot see. Keld Atlas first makes AI work legible — what it is, which workstream it serves, what it costs, what it returns — and optimization follows from that understanding. - **Two kinds of inference, one view.** Personal inference is people using AI tools (collected via Keld Signal, OpenTelemetry from tools such as Claude Code, Codex and Gemini CLI, and provider admin keys). Agentic inference is the company's own apps, agents and pipelines (collected via the Keld SDKs for Python and TypeScript and the Amazon Bedrock collector). Atlas reports both side by side, by team, person, tool, model and workstream. - **Inefficiencies, surfaced.** Atlas recommended actions flag idle subscriptions, members on the wrong plan, frontier models doing routine work, unmapped spend and teams heading over budget, ranked by impact. - **Workstreams organize the work.** A workstream is a declared line of work with an owner and a budget: a project, a marketing campaign, a customer effort, a GTM phase, an engineering repository. It is the only thing a customer defines. Every activity, dollar and outcome rolls up to it, across sessions, people, apps and agents. - **Activity types are the unit of work on the market.** Keld reads each prompt as the activities it actually contains — a step of work plus the context it is performed in — and classifies each activity into an **activity type** (for example: text creation, coding, extraction, transcription, Q&A). Activity types make inference comparable across customers and tradable on the marketplace. Routing rules filter by them, order books are based on them, quality is measured against them. - **Send any activity type to the Keld Marketplace.** Companies pick the activity types worth moving, from their apps and agents or their people's tools, and send them to the Keld Marketplace; everything else keeps running where it is. Batch and deadline-based processing are not part of the current offering. - Open-weight first. The marketplace gives the most visibility to popular open-weight models (Llama, DeepSeek, Qwen, Mixtral, Gemma, Phi and more) and the providers that serve them, alongside frontier models where the work genuinely needs them. - Privacy is architectural. Classification happens on the customer's own machines via Keld Signal; only labels, features and hashes reach Keld, never prompt or response content. - Key outcomes: meaningful cost reduction on work that does not need a premium model (internal estimates vary by workload and activity type); full attribution by workstream, team, model and activity type; real-time spend visibility, starting free. - Three founding principles. **Neutrality:** the platform favors neither side; it matches on price, measured quality and fit for the activity type, never on who pays Keld. **Transparency:** real prices and fulfillment volumes are visible, and every routing decision records route, price paid, baseline and savings. **Discoverability:** providers delivering near-frontier quality at a fraction of the cost get found and used, instead of being overlooked. --- ## How it works, the unit of work 1. **Declare a workstream.** The customer defines the lines of work that matter (project, campaign, customer effort, GTM phase, repository), with an owner, a budget and a period. Optionally imported from Notion, Monday.com or Jira. Nothing else has to be set up or tagged. 2. **Keep prompting as usual.** No behaviour change, no code changes. Keld Signal observes AI work on the customer's own machines and infers the rest. 3. **The work is typed.** Keld reads each prompt as the activities it contains and classifies every activity into an **activity type**: what is being done, carrying the context it is being done in. "Text creation — legal" is a different activity type from "Text creation — marketing", and they can route to different models at different prices. 4. **Routing.** Customer-declared rules pick slices of work by activity type ("all short text creation in this workstream") within a policy envelope: ceiling price, quality floor, compliance (privacy/no-training, geography, SOC 2, HIPAA), and named-model pinning where required. Each activity is routed individually and the decision is recorded and auditable. 5. **Clearing and execution.** Marketplace-addressable activities clear against supplier orders on an order book — either a Keld-curated activity-type book or an individual model book — and run on the winning provider. Work that is not marketplace-addressable routes through Keld or bring-your-own provider accounts at negotiated rates, with the same attribution. Most volume starts here. 6. **Attribution back to the workstream.** Cost, tokens, model, activity type, team and workstream are recorded for every activity, whichever path it took, and roll up into workstream outcomes: what the workstream returned against what it cost. Note on terminology: the concept of **Job Category** is retired and superseded by **Activity Type**. Anywhere older Keld material refers to job categories, activity types are the current model. --- ## Core Products ### Keld Atlas, for Enterprises (primary, free to start) * The control room where a company understands how it uses AI: personal inference (people in AI tools) and agentic inference (its own apps and agents) in one view, every activity attributed to a workstream, a team, a person, a tool, a model and an activity type. * **Inefficiencies:** recommended actions flag idle seats, members on the wrong subscription tier, frontier models on routine work, unmapped spend and budget overruns, each with its dollar impact. * **Usage and cost tracking for leaders:** adoption by team and person, spend against budget, Capex/Opex to accounting standards, scheduled email reports and CSV exports for ERP and BI. * **Workstreams:** declare projects, marketing campaigns, customer efforts, GTM phases and engineering repositories; Atlas attributes every activity, dollar and outcome to them automatically, with no tagging and no behaviour change. * **Spend observability:** ingests multi-source telemetry via API keys, OpenTelemetry and SDKs, and reports AI cost across workstream, team, model, provider, activity type and Capex/Opex, free to start. * **Outcome-based ROI:** cost per resolved ticket, closed deal, shipped feature or document processed, so spend is measured against the value it creates rather than tokens in isolation. * **Marketplace readiness:** Atlas shows which activity types in each workstream could move to the Keld Marketplace and what that would save, so sending them to the Keld Marketplace is a switch, not a rebuild. * Manage credentials and keys, set budgets and quotas at org, team and workstream level, configure policy. Self-hosted and air-gapped Atlas available. * **Use case:** A finance lead sees the "Customer onboarding" workstream costs $12,400/mo, that 46% of it is short text creation and extraction addressable on the marketplace, and turns on routing for those activity types. ### Keld Signal, on-device classification * Classifies AI work at the edge, on the customer's own machines, and ships **labels only** to Atlas: activity type, workstream, model, token count, timestamp. Prompt and response content physically never leaves the customer's environment. * The current release aggregates over **entire conversations** rather than isolated turns, which materially improves the accuracy of task, activity and activity-type detection. * Installed as a local hook for workstation tools (Claude Code, Codex, Gemini CLI, desktop assistants) or as an SDK / OTLP exporter for apps, agents and pipelines. * Zero Data Retention, encrypted in transit. Available for macOS, Windows and Linux. ### Keld Marketplace, for Enterprises (private beta) * Where enterprises buy AI inference, and where Atlas understanding turns into savings: companies send the activity types they choose to the marketplace, where specialist providers compete to run them and each clears on the best-value provider of the model that fits, within the customer's compliance and price bounds. * **Send work by activity type.** Rules select activity types ("all extraction in the support agent", "short text creation in this workstream") from apps, agents or people's tools. No code changes: Atlas routes centrally through the Keld SDK and Signal, and the Keld Marketplace API is an OpenAI-compatible endpoint for any client. * Every marketplace call records route, price paid, baseline and savings, so savings claims are auditable. Named-model pinning is always available for regulated buyers. * Trusted and secure: Zero Data Retention, SOC 2 (in progress) and GDPR, jurisdiction and model-origin bounds, and a host-your-own option. ### Integrations, for Enterprises * Drop-in integrations that speak IXP for the tools teams already use: OpenAI SDK, LangChain, LiteLLM, Vercel AI SDK, LlamaIndex, n8n, Zapier, plus native Python and TypeScript SDKs. * Measure now, optimize later, a switch you flip, not a rebuild. ### Keld Trade, for AI Model Providers (primary, private beta) * A trading platform with an SDK, UI and autonomous agent for providers to place, manage, reprice and cancel **orders** — offers to run work at a price, on a named model or on an activity type. * Orders rest in order books of exactly two kinds: a **Keld-curated activity-type book** (for example, models above a quality threshold for text creation) or an **individual model book**. Execution results feed back to re-rank and curate the books, so measured quality decides the flow. * Serves **live enterprise demand**: activity types companies send to the Keld Marketplace from their apps, agents and people's AI tools, delivered straight to the provider's endpoints. * Runs algorithmic yield strategies (e.g. MAX_YIELD) based on live demand, with automated clearing, settlement and a full audit trail. * **Use case:** A GPU operator posts asks on two activity-type books, and fills idle headroom with enterprise demand sent to the Keld Marketplace. --- ## Key Benefits Summary * **For enterprises:** understand how the business uses AI across people and agents, what it costs and what it returns; find the inefficiencies; stop overpaying for premium models the work never needed; full attribution by workstream, team, model and activity type; auditable savings; no lock-in and no code changes. * **For AI model providers:** turn idle capacity into revenue; sell on named models or on activity types; live enterprise demand; automated order, yield and settlement management. * **For both sides:** a neutral, transparent marketplace where typed activities make inference comparable, prices are visible, and settlement is auditable. * **Open-weight first:** the marketplace surfaces popular open-weight models and providers, so high-quality, low-cost options get discovered and used. * **No migration:** every integration implements IXP, so measuring, and later optimizing, is a switch you flip, not a rebuild. --- ## Keld vs. the status quo #### Direct, fixed-price API contracts * Status quo: you pay list price for a premium model whether or not the work needs it. * Keld: each activity clears at market price within your ceiling; you pay only for the intelligence that piece of work requires. #### Manual multi-provider routing * Status quo: engineering maintains bespoke routing logic per provider, on anonymous requests. * Keld: work is typed before it is routed, so rules are written in business terms ("short text creation in this workstream") and matched on the best-value provider that fits. #### Gateways and routers * Status quo: a router optimizes a stream of anonymous prompts and has no idea what the work means or what it was for. * Keld: activities carry their context and their workstream, so routing, reporting and ROI all speak the same language. #### Opaque pricing * Status quo: pricing and utilization are black boxes. * Keld: real prices and volumes are visible, every routing decision stores route, price paid, baseline and savings, and Atlas attributes spend across every vector. --- ## Plans ### Understand (spend observability), Free * Personal and agentic inference in one view, recommended actions on inefficiencies, workstreams, activity-type classification and full inference spend attribution across workstreams, teams, models, activity types, Capex/Opex and outcome-based ROI. No code changes beyond a drop-in SDK, the Signal hook or an OpenTelemetry config. ### Optimize (Keld Marketplace), Usage-based * Route the activity types you choose to the best model at the best price, within your ceiling and compliance bounds. Pay on optimized volume. ### Enterprise * SSO, granular roles, private marketplaces, self-hosted Atlas, named-model enforcement, SOC 2 (in progress) / GDPR compliance, and dedicated support. > Access is currently via private beta, see Contact. --- ## Protocol (developers) - [IXP, Inference eXchange Protocol](https://keld.co/docs): the open wire standard for pricing, matching, asset reservations, execution tokens and settlement data. A developer-facing technical topic. Implement it once (via Keld SDKs, integrations, or a custom adapter) and your work can run anywhere on the marketplace. Keld cleanly separates the control plane (financial settlement) from the data plane (live execution). --- ## Typical data flow 1. A customer declares a workstream in Atlas (owner, budget, period, routing rules). 2. People and agents keep working as they always have; Keld Signal classifies activities on-device and sends labels only. 3. Atlas attributes each activity to its workstream, team, model and activity type, and reports spend, budgets and ROI. 4. Routing rules select activity types; each one sent to the Keld Marketplace clears on the best-value provider within the customer's bounds. 5. Marketplace-addressable activities clear against supplier orders on an activity-type or individual model book; the rest run on Keld or bring-your-own provider accounts. 6. Results return to the customer; the routing decision, price paid, baseline and savings are recorded, and outcomes roll up to the workstream. --- ## Implementation checklist **For enterprises** 1. Create a free Atlas account and declare your first workstreams (projects, campaigns, customer efforts, GTM phases, repositories). 2. Install Keld Signal for workstations, or the SDK / OpenTelemetry exporter for apps and agents. Spend and activity types appear in minutes. 3. Review which activity types dominate each workstream, how much is marketplace-addressable, and the projected savings. 4. Send those activity types to the Keld Marketplace; set ceiling prices and compliance bounds, with no code changes. 5. (Optional) Use Atlas for budgets, quotas and policy at org, team and workstream level, or add Integrations to your existing stack. **For AI model providers** 1. Connect Keld Trade to place and manage orders on named models or activity types. 2. Choose or build an order strategy (e.g. MAX_YIELD). 3. Earn on idle capacity, with automated clearing, settlement and audit trail. --- ## Resources - [Keld Atlas](https://keld.co/atlas): see how your people and agents use AI and what it costs, find inefficiencies, free to start. - [Keld Marketplace](https://keld.co/marketplace): buy AI inference from providers competing on every activity type, private beta. - [Keld Trade](https://keld.co/trade): the trading platform for AI model providers (private beta). - [Documentation](https://keld.co/docs): guides and references across the platform, Getting Started, Keld Atlas, Keld Signal, and Integrations & SDKs. --- ## How to engage Most enterprises start free with Atlas: declare a few workstreams, install Signal, and watch AI work become legible by activity type. Once that data exists, the activity types worth moving go to the Keld Marketplace and begin saving. AI model providers start with Keld Trade to monetize spare capacity. ## Contact - [Join the Atlas private beta](https://keld.co/signup): the current way to get access. - [Contact](https://keld.co/contact): questions, walkthroughs, supplying models, partnerships. - Email: founders@keld.co - X / Twitter: https://x.com/keld_exchange - LinkedIn: https://www.linkedin.com/company/keldex ## References - [Keld website](https://keld.co/), the public landing page describing the platform and how it works. - [Keld Atlas, for enterprises](https://keld.co/atlas) - [Keld Marketplace, for enterprises](https://keld.co/marketplace) - [Keld Trade, for AI model providers](https://keld.co/trade) - [Documentation](https://keld.co/docs) ## Optional Secondary context, safe to skip if your context budget is limited. Blog posts are listed newest first. - [Introducing Keld: a marketplace for AI work](https://keld.co/blog/introducing-keld): the company announcement by Dave Otten, Doug Shore, Federico Enni and the Keld team. Why Keld exists (companies can't tie AI spend to outcomes or forecast it, and inference needs a neutral marketplace at the inference layer), the mission, and how Keld Atlas and the Keld Marketplace work together. - [Workstreams, a new way to report on your AI](https://keld.co/blog/introducing-workstreams-in-atlas): why a token tells you what was consumed but not what was accomplished, and how a Workstream rolls the tasks and activities under a piece of business work into one number. - [Introducing Keld Atlas](https://keld.co/blog/introducing-keld-atlas): the launch of Keld Atlas in private beta; why Keld starts with understanding AI use before optimizing it, what Atlas shows across personal and agentic inference, and how Keld Signal keeps prompts on device. - [Keld vs OpenRouter and LiteLLM](https://keld.co/blog/router-gateway-vs-marketplace): how Keld differs from LLM gateways; it maps company-wide AI use without reading prompts, optimizes models, subscriptions and routing, and prices inference on a neutral two-sided marketplace instead of list price plus a fee. - [Why an inference marketplace](https://keld.co/blog/why-an-inference-marketplace): essay on why the same inference can cost 5 dollars or 5 cents, how much that fragmentation makes buyers overpay, and how a neutral marketplace that prices each activity type (Keld Atlas, the Keld Marketplace and Keld Trade) fixes it. - [IXP, Inference eXchange Protocol](https://keld.co/docs): the open wire standard for pricing, matching, asset reservations, execution tokens and settlement. - [Security](https://keld.co/security): control-plane / data-plane split, on-device classification, Zero Data Retention, subprocessors and compliance status. ## Legal & trust - [Trust Center](https://keld.co/security): hub for security, data flow & residency, certifications (SOC 2 and ISO 27001 in progress) and trust documents. - [Privacy Notice](https://keld.co/privacy): how Keld handles personal data; controller (website/account) vs processor (inference payload) roles; GDPR and CCPA/CPRA rights. - [Cookie Policy](https://keld.co/cookies): every cookie by name, purpose, duration and provider; consent-first. - [Terms of Service](https://keld.co/terms): the agreement governing use of Keld. - [Acceptable Use Policy](https://keld.co/aup): what's allowed on the marketplace and who is responsible for submitted data. - [Sub-processors](https://keld.co/subprocessors): published list of infrastructure vendors and supply-side providers.