llms.txt
+21 −38
− > Saina is the decision layer for software systems. Saina builds specialist AI decision models and the developer platform around them. Developers use Saina to turn model predictions into versioned decision endpoints, observe how those decisions perform in production, turn failures into test cases, and improve their models or policies without breaking existing behavior. Saina works with open models and external inference providers, so teams can use their own infrastructure.− Saina is operated by Rama Labs Inc. Helm is the family of specialist decision models. Saina Studio is the developer platform for building, testing, deploying and improving decisions. Applications supply context, questions and candidate answers; Helm returns structured scores rather than free-form generated answers, and the application controls what happens next. Several questions can share the same context in one request. Typical decisions are routing, gating, assignment, allocation, scheduling,− In decision mode, thresholds and minimum margins determine whether an answer is accepted. Low scores, a tie or an insufficient gap can produce an explicit fallback. The application decides what happens next: ask a person, gather more context or call another model. A probability score is not a guarantee of correctness or evidence of calibration on a particular workflow; evaluate the model and policy on your own held-out examples.− ## Models and answer types− - **Helm 3** (`saina-helm-3-0.8b`): a native multimodal decision model that reads text, images and audio together.− - **Helm 2** (`saina-helm-2-0.8b`): text decisions, including classification and routing.− - **Helm 2 Omni** (`saina-helm-2-omni-0.8b`): text, image and audio decisions; remains available alongside Helm 3.− - **Helm 2 Routing** (`saina-helm-2-routing-0.8b`): text-based selection of an answering model using the request and candidate capabilities and prices.− - **Mechanical specialist**: use `specialist: "mechanical"` with Helm 2 Omni for machine-sound decisions. This specialist accepts text and audio, not images.− Supported question types are `single_choice` (mutually exclusive options), `yes_no`, `rating` (an ordered rubric) and `multi_choice` (independent scores for overlapping labels). Multi-choice scores do not need to sum to one. See the API documentation for each type's request and response fields.− Helm weights are published under Apache-2.0. Check the selected model card for capabilities and limitations, and pin a checkpoint revision for reproducible self-hosting. Hosted availability is reported by `GET https://api.saina.run/v1/models`; the `saina-helm` alias refers to the model served by the endpoint, rather than a fixed release.− ## Saina Studio− Studio turns a model into a production decision endpoint and keeps it reliable: a playground for trying models, inputs and questions; versioned decision endpoints served through the API; observability that records decisions, scores, execution traces and reported outcomes; test cases with expected results; regression evaluation that compares versions; promotion of a tested version to production and rollback to an earlier one; and data labeling and fine-tuning that turn observed failures into trai− ## Integration and deployment− Use the hosted API at `https://api.saina.run`, or run Helm locally with the Saina SDK or a self-hosted HTTP server. Python and JavaScript clients and an n8n integration are available. CPU and GPU deployment instructions are in the deployment repository; latency and memory depend on the model, runtime and input. Use the current API pages for pricing, account access, limits and data handling.− ## Studio MCP− Saina Studio exposes an MCP server at `https://api.saina.run/v1/mcp` (HTTP transport). It gives an AI client, such as Claude, Claude Code, Codex or Cursor, access to one account's Studio tools, with the access level chosen when the connection is approved. Clients with sign-in support connect through OAuth: add the URL as a custom connector, sign in to Saina Studio, choose Read only or Read and write, and approve. No key is copied, access renews for up to 90 days, and the account owner can revoke− ## Product and documentation− - [Saina Helm](https://saina.run/): Model family, benchmarks and deployment options.− - [Decision models](https://saina.run/decision-model/): What a decision model does and when to use one.− - [Hosted API](https://saina.run/api/): Hosted service overview.− - [API documentation](https://saina.run/api/docs/): Request formats and integration guidance.− - [Playground](https://saina.run/studio/): Try model inputs and inspect scores and decisions.− - [Agent skill](https://saina.run/skill.md): Connect to Saina Studio over MCP, the tool groups, common workflows and rules for agents.− - [Studio MCP](https://saina.run/studio/): Endpoint, sign-in and API-key setup for MCP clients (Studio, Settings, MCP).− - [Saina SDK](https://github.com/run-saina/saina): Python and JavaScript clients, decision contract and optional local inference.− - [Self-hosting](https://github.com/run-saina/deploy): CPU and GPU deployment instructions and container images.− - [n8n integration](https://github.com/run-saina/n8n-nodes): Workflow integration.− - [Model weights](https://huggingface.co/run-saina): Published Saina models and model cards.− - [Decision Index](https://github.com/run-saina/saina-decision-index): Evaluation harness and methodology.− ## Use cases− - [Use cases](https://saina.run/use-cases/): Workflow examples.− - [Support ticket triage](https://saina.run/use-cases/support-ticket-triage/): Classify and route tickets.− - [Email inbox routing](https://saina.run/use-cases/email-inbox-routing/): Route incoming messages.− - [LLM routing](https://saina.run/use-cases/llm-routing/): Select models and gate actions.− - [Document and receipt checks](https://saina.run/use-cases/document-receipt-checks/): Compare documents with claims.− - [Machine fault detection](https://saina.run/use-cases/machine-fault-detection/): Classify equipment sounds.− - [Blog](https://saina.run/blog/): Product announcements and integration articles.+ > Saina checks what your agents did against the policies and SOPs they should follow, using the OpenTelemetry traces your systems already record, and shows what went wrong and the evidence behind each issue.+ Saina is operated by Rama Labs Inc. Use Saina Studio in your browser or through a compatible AI assistant.+ ## What Saina helps with+ - Turn a rule from a policy, SOP or agent instruction into a check, and try it on recent runs before monitoring.+ - Flag runs where approval was missing, a deadline passed or a handoff never happened.+ - Connect what the agent should have done, what it did, and the trace that explains the difference.+ - Report a likely cause only when the trace supports one; otherwise show the gap and the missing evidence.+ - See when the same step keeps being missed across runs, and get Slack alerts naming the run, the missed requirement and the trace to inspect.+ ## Getting started+ Describe an obligation in your own words — for example, "the agent must get manager approval before issuing refunds over $500" — including any deadlines or exceptions. Send OpenTelemetry traces from the systems your agents run on, try the check on recent runs, and choose when to be alerted. The same checks also cover work done by people and automated workflows.+ Saina accepts OTLP/HTTP traces in protobuf or JSON. The traces need to record the agent's messages, tool calls and actions, plus the outcomes the rule depends on. Results show when they were last checked and link to the traces behind them. Runs with no relevant trace are reported as missing evidence rather than silently passed; review those before deciding a rule was broken.+ ## Pricing+ Credits pay for AI checks that read agent runs. The amount depends on how much of each run a check reads; checks based only on fields or deadlines use no credits. Studio estimates the cost before paid checks start. Paid plans include monthly credits. Plans: Free (1 credit on each day used, kept 30 days), Starter $5 a month (25,000 credits), Pro $10 a month (55,000 credits), Max $100 a month (625,000 credits). Details: https://saina.run/pricing/+ ## MCP+ Saina Studio exposes an MCP server at `https://api.saina.run/v1/mcp` (HTTP transport). Clients with sign-in support connect through OAuth: add the URL as a custom connector, sign in to Saina Studio, choose Read only or Read and write, and approve. Clients without sign-in send an API key with the `mcp` scope in the `Authorization: Bearer` header. For example, `claude mcp add --transport http saina https://api.saina.run/v1/mcp` connects Claude Code with sign-in. Per-client setup is at `https://sai+ ## Pages+ - [Saina](https://saina.run/): What Saina does and how it works.+ - [Pricing](https://saina.run/pricing/): Plans and credits.+ - [Agent skill](https://saina.run/skill.md): Connect to Saina Studio over MCP and the rules for agents.+ - [Saina Studio](https://saina.run/studio/): Sign in, send OpenTelemetry traces and set up checks for your agents.+ - [Data handling](https://saina.run/api/#data): How Saina-hosted model requests are handled.