Who it's built for
The regulated team
Compliance is the brake. Spegling is the brake-release.
Finance, legal, healthcare, regulated SaaS. Your compliance team blocks AI adoption because no one can answer "which prompt produced this decision, on which model, with what rollback plan." Spegling is the audit substrate beneath whatever agent stack you pick. Pilot it under your hardest scenario.
Finance · Legal · Healthcare · Regulated SaaS
The multi-hat operator
One human, three fiduciary hats, three authority ceilings.
You're a founder, advisor, and board member, often in the same week. The same AI agent helps with all of them. Spegling binds authority to the role, not to the human. Board work carries fiduciary duty to one company; advisory work carries duty to another. The agent knows which.
Solo founders · Advisors · Board members
The team starting right
Arrive at compliance with the architecture.
You're standing up AI inside the company and want the audit trail, the role registry, and the rollback plan built in from day one. Skip the cycle of "ship fast, audit later." The same governance you'd be forced to bolt on after the first incident, but in from the start.
Mid-market · AI-led startups · Engineering teams
How you use it
- From the tools you already use. Spegling speaks MCP. Claude Desktop, Cursor, your team's own agent. Anything MCP-aware plugs in and reaches durable memory, intent, evidence, and the corpus. Governed capabilities live inside Spegling. The external agent gets to use them; it cannot sidestep them, because there is no second path to the same data.
- Bring your own agent. Any provider. Run it inside a per-tenant container with persistent state, an encrypted vault for provider keys, and an egress firewall that opens only to the providers you allowlisted. No infra to operate; what you bring is the agent code.
- Long-running work the governor holds. Multi-hour and multi-day runs are first-class. The governor schedules, checkpoints, enforces the budget, and surfaces decisions for human approval. Comes back when the task does.
What you build with it
- Authority that names whose hat. The same operator wears three hats: board member, advisor, investor. Three different authority ceilings, defined once at the tenant. Every delegation checks against the role before the grant lands. The schema enforces it.
- Action chains over typed channels. Handoffs, ledger writes, and approvals all flow through structured tool calls. A document landing in your wiki cannot pretend to be a control signal.
- Embedding egress that strips secrets at the wire. One redaction chokepoint catches PEM blocks, provider API keys, AWS access ids, GitHub tokens, JWTs, Bearer headers, and env-var assignments. Every embed call benefits without per-site wiring.
- A ledger that names the prompt and the model. Every prompt template carries an immutable version id. Every inference record carries the version. "What produced this decision, on which model, at what cost, with what rollback plan" is a single look-up, not a forensic exercise.
Decision audit · 2026-06-17 · 23:42 UTC
- Intent
- Apply wiki rule "require-citations"
- Prompt
- chat_system v4 (immutable)
- Model
- claude-sonnet-4
- Cost
- €0.18 over 4.2s, 12,400 input tokens
- Approved by
- hannu@varjosoft.com
- Rollback
- Restore prior body "Every claim must cite a source"
- Self-improvement with a real apply step. Proposals from agents land inert. A fresh-context skeptic reviews each one without seeing the proposer's framing and writes a verdict you read. Promotion requires a human user and a written rollback plan. The schema rejects every other path.
How it's built
Five primitives. Every link in the chain hangs off the intent id, so authority, evidence, and rollback travel together by design.
Declared Intent
Stable record carrying scope, constraints, and the operator's role.
Living Authority
Ceiling enforced at delegation creation. Re-asked when the world moves underneath.
Evidence Audit
Stamped with prompt version, model, cost, and latency. Indexed by intent. Queryable across runs.
Inert Proposals
Self-improvement gated at the schema. An adversarial review attaches before any human is asked.
Durable Memory
Anything written traces to a verified tool-call result. Reasoning chains live in working memory.
