ObVectinfrastructureallegro
Tune
Each step lands on the model that fits its cost, quality, context, and latency budget — and the routing decision is logged, so you can audit why a provider won.
Op. 1 · the AI runtime
One API.
Every model, in harmony.
A multitude of providers — each with its own interface, limits, and quirks — is noise. ObVect resolves it into one score: inference routed per step, answers grounded in your corpora, and multi-agent work that keeps a single thread.
BYOKbring your keysREST · MCPone interfaceReplayableevery run
Run in G — a versioned reasoning graph
conducted by ObVecttempo: cost-cappedgate: write_pr awaits approvalmanifest: saved
IOverture
From a multitude of providers to one system.
Three principles carry every request — the reason the name begins with a circle seen from above: many voices, one centre.
andante
Structure
Any model can run as a multi-turn agent. Plans, tool calls, and approvals live on one versioned graph — saved as a manifest you can replay.
tutti
Harmony
One shared context follows the work. Hand a task between agents and models without re-priming, and keep answers grounded in tenant-scoped sources.
IIThe score · Virtuo SDK
The reasoning-graph and routing engine.
Virtuo is ObVect's model-agnostic engine for multi-agent orchestration: every model gets agent behaviour under policies you control, with a human at the moments that matter.
- agent loop
- Models with native multi-turn run as-is; every other model gets the same plan → act → observe → iterate loop.
- shared map
- Agents read and write one context of plans, decisions, and results — explainable, auditable, and portable across models.
- approvals
- Bound to your PM or reasoning system: the right role signs off, and policy decides what may run unattended.
- model routing
- A built-in gateway — a model catalogue with tiers, shared keys or full BYOK/BYOM, and routing that keeps learning from real outcomes.
- decision model
- A small open-source router resolves routes and approvals in milliseconds with calibrated confidence — and escalates when unsure.
- machine identity
- Every agent runs as its own Idio-issued identity; outbound calls carry a signed agent profile a third party can verify.
Cue sheet — one run
graphplan → act → observe → verify
routecost_capped · fallback: local
contexttenant/acme · 6 sources cited
toolsrepo.search, tests.run
approvewrite_pr, deploy
identityreviewer@acme · revocable
replaymanifest saved
Example manifest. Every run is versioned and replayable.
IIIThe ensemble
Local, self-hosted, or every major provider.
Seat the models you already trust. ObVect conducts; you choose the players.
- iLocalA model on your machine — private and offline-capable.
- iiBring your keysAnthropic, OpenAI, Google, and the rest, routed per request.
- iiiHosted endpointsYour own compute without owning hardware.
- ivTenant modelA private model for your tenant, trained from approved usage. Never shared.
Any OpenAI-compatible endpoint — including a self-hosted GPU deployment — can register as a custom provider; privacy mode keeps prompts off third-party routes.
—Rehearsal · a run
Plan, route, act, verify.
1
Plan
The graph decomposes the request into scoped steps.
2
Route
Each step lands on the model and tools its policy allows.
3
Act
Agents execute; consequential actions pause for approval.
4
Verify
A second pass checks the result before it is returned and recorded.
IVComposition · Ladder SDK
Terraform for products — from gaps to shipped code.
Ladder is the product-graph compiler that runs on ObVect. It maps connected repos, finds what is missing, and dispatches the work until the graph closes.
| Map | A product graph across every connected repo — missing pieces, duplicates, coverage holes, orphaned work. |
| Analyze | Gaps become tickets, deduped against your boards and sequenced in dependency order. |
| Dispatch | Each ticket routes to the right agent for its layer, under a scoped identity. |
| Verify | Every cycle re-checks what landed; aim a pass at ship readiness, bugs, docs, or security. |
—The reference library · Corpus MCP
The substrate, callable from anywhere.
ObVect's analysis draws on a curated corpus — exemplar code, docs, Stack Overflow evidence, and connectors across 145 domains. The Corpus MCP exposes it to any MCP client: routing, search, and citations over one endpoint.
∿
Route
corpus_route — "what should I build from" answered with a scored, licensed exemplar, not a guess.
∿
Browse
corpus_search + corpus_item — the lane index and full evidence records, redacted of tenant internals.
∿
Learn
corpus_outcome — outcome reports drift routing quality and calibrate to your tenant.
∿
Gate
Anonymous callers get a teaser rate; an account key unlocks the sustained tier. OAuth scopes or a shared key — your choice.
—The pit, not the stage
Infrastructure, not a tenant product.
ObVect has no workspace of its own. It plays inside the entity context Idio resolves and the scopes each product defines — and serves the AI behind all of them.
Identity
Idio
Supplies the user, entity, and agent identity for every call.
Execution
ObVect
Routes, retrieves, and runs the workflow under that policy.
Record
Your product
Receives outputs, traces, and costs for its own audit log.
VProgramme notes · governance
The controls your security team will ask for.
The same identity, policy, provenance, and consent controls run under every Colossal Capital product.
- AI provenanceEvery AI-assisted result in ObVect is disclosed and traceable — which persona acted, which model ran, on whose authority, from which sources.
- Built for AI regulationTransparency notices, human oversight, risk documentation, and audit logs designed to support EU AI Act, Colorado AI Act, and NIST AI RMF / ISO/IEC 42001 programs.
- Private by designEntity-scoped storage, data minimization, encryption in transit and at rest, and GDPR / CCPA-CPRA request handling. Customer data never trains a shared model without opt-in.
- Fine-grained provisioningIdio SSO over SAML and OIDC, entity context, role and attribute policy, and a named, scoped, revocable identity for every agent.
- Humans at the gateConsequential actions wait for an eligible approver, and every decision lands in an exportable audit trail.
Private models, at the scope you choose — nothing rolls up on its own.
- UserOne user's approved prompts and feedback
- ContentOne workflow, tool, or granted corpus
- EntityOne tenant's private adapter
- OrganizationAn organization model from permitted corpora
Every run manifest records the persona, model, tools, sources, and approvals involved, so any output can be traced and replayed. Shared ecosystem learning is a separate, sanitized, opt-in path. AI transparency notice.
Finale
Build on the runtime the products run on.
Request API access for the beta and route your first workflow in minutes.