AI agents your team can trust, inspect and measure.
AgentForge answers from your documents with citations, uses tools only through explicit capabilities, asks a person before it changes anything, and records every step so you can see exactly why it did what it did.
The demo signs you in to a seeded workspace. It runs on a deterministic demo model, so it works without any API key; connect Claude by setting ANTHROPIC_API_KEY.
Grounded answers with citations
Upload Markdown, text or PDF. Documents are chunked, embedded and searched with hybrid retrieval; every answer cites the passages it used.
A real agent loop
Classify → plan → retrieve → call tools → validate → answer, as a checkpointed LangGraph workflow that survives restarts and retries.
Humans approve side effects
Tools declare explicit capabilities. Anything that writes data waits for a person with the right role to approve or reject it.
Run inspector
Every step, model call, token count, tool input and output, approval and error of every run, streamed live as it happens.
Evaluations and regressions
Datasets of expected behaviour run against pinned agent versions, scored by evaluators, compared with the previous run.
Teams, roles and audit
Organizations, projects, four roles, tenant isolation in every query, and an append-only audit log of who did what.