arcp is the operator's seat for agent companies. Register your fleet, wire it to sandbox, gateway, and inference providers, and watch every run unfold over a single API and UI. Built on Tensile.
Every serious agent shop ends up rebuilding the same plumbing: identity and orgs, an agent registry, dropdowns for sandbox / gateway / inference, a run timeline, observability. arcp is that layer — multi-tenant, OIDC-first, with Tensile wired in as the default substrate.
Orgs, roles, API tokens, OIDC SSO — the same pattern Tensile uses. Bring your IdP, give your engineers their own keys, and isolate every run by tenant.
An agent declares the capabilities it serves. Operators register their fleet once and reuse them across runs. Capabilities are still job kinds — the data model stays clean.
Sandbox, gateway, and inference default to Tensile. Swap in Modal, OpenAI, Anthropic, or your own provider per agent — without touching the control-plane code.
Four production agents ship in every arcp install. Each one is
an AgentEntry you can clone, fork, or swap out for
your own. Sign in to inspect their configs and runs.
Refactors, fixes bugs, and applies research recipes against
a repo. Capabilities: code_improvement,
bug_sweep, repo_comprehend.
Converts a paper into a working implementation. Sibling
HTTP service, six stages, human-in-the-loop gates.
Capability: paper_to_code.
One web-search and scraping agent reused across every
research workflow. Capabilities: research_sweep,
thesis_explore.
Sweeps over inference configurations to find optimal
cost / quality / latency points. Capabilities:
autoinfer, ab_test,
autospecdecode.
Three steps from zero to a live agent fleet. Each step is backed by a first-class page in the product.
Walk the 7-step wizard. Name, repo URL, capabilities,
transport, and resource bindings — all dropdowns,
persisted as an AgentEntry in the registry.
Choose your sandbox, gateway, inference, evals, rubrics, and verifiers from a single page. Tensile is preselected wherever it's the default.
Trigger runs from the agent detail page, the API, or your
CI. Every run lands in /runs with a live
event feed, aggregate metrics in /observability.
Eight features the rest of agent ops keeps re-implementing. They're all in arcp out of the box.
Orgs, roles, per-tenant agents, per-tenant resources, per-tenant runs.
Authentik, Okta, Azure AD, Auth0 — any OIDC provider. PKCE, JWKS caching, sessions + CSRF.
Firecracker microVMs and the Front Door API wired in. Modal and local execution are peer options.
Register your own via the wizard or the POST /v1/agents API. In-process, HTTP, or subprocess transports.
One polymorphic store for sandboxes, gateways, inference, models, evals, rubrics, verifiers, SDKs, finetuning.
Parent / child grouping, live event feeds, stages visualized — the whole run, one page.
Runs per day, cost per agent, error rate per kind. Server-side aggregation in SQL, no external metrics store needed.
arcp_ tokens, Argon2id-hashed, org-pinned, role-scoped. Browsers get cookies + CSRF.
arcp is one of the Research Commons stack of agent infrastructure products. They're designed to work together.
Firecracker microVMs + K8s control plane. Submit work, stream logs, get back artifacts.
Pluggable code-improvement agent. Repomap, code graph, embeddings, multiple tracks.
Six-stage pipeline that converts a paper into a working implementation, with HITL gates.
One web-search / scrape agent reused across every research workflow. Sibling HTTP service.
Research Commons runs every internal agent on arcp. We're opening it up to a small group of design partners.
Schedulers, sandboxes, observability, and shared services that every RC product runs on. arcp is one of those products.
Two repos, two terminals, two minutes. Sign in via OIDC, claim the bootstrap token to become an admin, then register your first agent.
# Clone + boot the control plane git clone git@github.com:Research-Commons/auto-research-control-plane.git cd auto-research-control-plane make serve # → JSON API + WebSocket on http://127.0.0.1:8765
# In a second terminal git clone git@github.com:Research-Commons/auto-research-control-plane-ui.git cd auto-research-control-plane-ui npm install && npm run dev # → open http://127.0.0.1:3000