Forest

The platform reads and writes directly to the customer's databases and APIs with no copy layer, with automatic schema introspection across Postgres, MongoDB, BigQuery, Snowflake and REST or GraphQL APIs. SDKs in any language handle coded actions and custom logic.

Workflows run on top of KYC and screening providers such as Sumsub, Onfido, Veriff, Persona and Trulioo, with every provider call recorded. An MCP client consumes any data source, and standard and custom MCP connectors come on every plan.

Agents take authenticated action in outside systems through record level actions and workflow steps, and the SDK adds custom tools.

SourceForest, forest.app and forest.app/pricingread 2026-09-06

Multi step workflows carry approvals, escalation and human in the loop gates, and case inboxes come with dispatch rules, segments and SLAs. Named processes such as KYC and KYB onboarding, regulator information requests, sanctions screening, transaction monitoring and dispute handling run as workflows where agents, automations and humans work the same case. The platform also has a workflow builder, approvals and escalations, LLM steps in workflows and a dispatcher for smart routing and assignment. Workflows are native multi step automation with no external stitching.

SourceForest, forest.app, forest.app/pricing and forest.app/blog/here-comes-the-new-forestread 2026-09-06

Agents ground on the customer's live records. For a regulator request, case context is assembled across CRM, KYC and transaction systems, drafts cite source records, and prior case matching handles repeat questions, all read directly from the customer's own datasources with no copy layer. This context is assembled per run over live operational data. There is no maintained retrieval structure, index, graph or embeddings layer over the customer's knowledge, and the data layer is the application's live database rather than a knowledge index.

SourceForest, forest.appread 2026-09-06

High risk actions pause for human approval before they run. Each action's risk and approval requirement is configured, routine actions matching configured patterns can execute without review, and the approval chain is the same for agent initiated and human initiated work. Two person review and per jurisdiction scoping apply to agents, and approvals and escalations come on every plan. Forest runs this review and approve surface itself, enforced at the data layer, rather than relying on a gate the customer already owns.

SourceForest, forest.app/platform/ai-governance-llm-ops and forest.appread 2026-09-06

Access control covers basic and advanced RBAC with custom roles, AI agent RBAC with scope and permissions, SSO and SAML from the Starter plan, SCIM provisioning and IP whitelist on Enterprise only, two factor authentication, and external user provisioning for partners and BPOs. Each agent connects under a role defining what it can read and trigger, enforced at the data layer. Forest names SOC 2, HIPAA and GDPR compliance without stating the SOC 2 type or auditor, and its Trust Center at trust.forest.app is a Vanta portal.

SourceForest, forest.app/pricing, forest.app/platform/ai-governance-llm-ops and forest.appread 2026-09-06

Every agent action gets a record level audit entry recording the model used, the agent's identity, the input data and prompt, the reasoning trace where the model exposes it, tool calls, the action taken and the outcome. Each entry is written at the moment of the action against the record it touched, in the same trail and format as human actions.

Regulator ready exports filter by agent, model, customer or time window and bundle the full action and reasoning trail, and events forward to Drata, Vanta, Sprinto, Secureframe, Datadog, Splunk or a warehouse by webhook or sync. AI agent audit trails and unlimited audit log retention come on every plan, with decision trace reasoning logs from Professional. Forest keeps this runtime trail itself to reconstruct why an agent acted, rather than relying on logs from the customer's systems.

SourceForest, forest.app/platform/ai-governance-llm-ops and forest.app/pricingread 2026-09-06

Case and record state persist in the customer's own database as the application's data model, and workflows hold their state across steps. There is no separate memory layer with its own scope and lifetime, and no per user, per team or per workflow memory. Forest does not say how the Forest Assist sidebar keeps session state.

SourceForest, forest.app, forest.app/pricing and forest.app/platform/ai-governance-llm-opsread 2026-09-06

Forest hosts its UI and runs a backend on the customer's own infrastructure that reads and writes the datasources directly, so data never leaves the customer's environment. Every plan, including the free Sandbox, comes with a self hosted Forest Agent. The backend, business logic, permissions and audit trail are deployed in the customer's infrastructure, client hosted, and audit logs and decision traces stay in the customer's environment rather than the vendor's. The vendor hosted UI is a control surface and does not change where the data resides, and no region is given for it.

SourceForest, forest.app/pricing and forest.appread 2026-09-06

Named process workflows cover KYC and KYB onboarding, regulator information requests, sanctions screening, transaction monitoring and dispute handling, each a configurable workflow on top of the providers the customer already uses. After mapping the customer's ecosystem, Forest's forward deployment team designs three to four workflows with the customer and takes them live over twelve weeks.

These are patterns the customer assembles with Forest rather than packaged agents or templates a customer adopts as they stand, and there is no template library or catalog of adoptable agents. The ecosystem's AI Agents listing names third party agents such as Sphinx, Diligent.ai and Rulebase that run under Forest's governance, which makes it an integration catalog.

SourceForest, forest.app and forest.app/platform/ai-governance-llm-opsread 2026-09-06

Automated triggers on time, data and webhook come from the Starter plan. A dispatcher handles smart routing and assignment alongside inbox and case management, and case inboxes carry dispatch rules, segments and SLAs. Intake arrives from a shared inbox, a partner portal or manual entry with deadlines attached, agents call the same workflows through the MCP server, and Forest can be embedded inside Zendesk, n8n and Claude. Work therefore reaches the agent through scheduled, data change and webhook events, an inbound queue with dispatch rules and several inbound channels.

SourceForest, forest.app/pricing and forest.appread 2026-09-06

The customer brings Claude, GPT or open source models, can switch providers without changing the audit shape or permission model, and has the model name recorded with each action. Anthropic, OpenAI, Mistral and Google Gemini power Forest Assist and the customer's agents, and bring your own model is part of what Forest calls open abstraction. Inference runs against the customer's approved model rather than one the vendor selects.

SourceForest, forest.app/platform/ai-governance-llm-ops and forest.appread 2026-09-06

Every plan includes API access for configuration and data, native SDK support (Node.js, Rails and others) for the self hosted Forest backend, and an MCP Server that exposes Forest to AI agents. The MCP server exposes data, schemas, actions and workflows to any authorized runtime, from Claude to open source models or custom built agents, so Claude and other agents can call Forest under the customer's governance. It runs inbound, letting an outside caller drive the platform, while a separate MCP client consumes outside data sources.

SourceForest, forest.app/pricing, forest.app and forest.app/blog/here-comes-the-new-forestread 2026-09-06

Workflow analytics and optimization come from the Professional plan, giving a team a live metric on operations to read when tuning a workflow, and the decision trace reasoning log lets a reviewer see how an agent reached a decision after the fact. Forest offers no evaluation harness, scored test cases, release gate or controlled loop for measuring a change to the agent. It says it does not evaluate model output for accuracy or run safety tests, and it points customers to upstream tools.

SourceForest, forest.app/pricing and forest.app/platform/ai-governance-llm-opsread 2026-09-06

Forest's own agents and workflow steps reach outside systems through direct database access, APIs and MCP connectors, and the platform drives no browser, desktop or computer control. The ecosystem lists Sphinx, a partner's browser native agent that logs into compliance tools like an analyst, but it is an outside agent that calls Forest under its governance rather than a browser engine Forest operates.

SourceForest, forest.app/platform/ai-governance-llm-ops and forest.appread 2026-09-06