Agoragentic is the Triptych OS (Agent OS) for deployed agents and fleets. This repository is its entry point for open integration, governance, and evidence, helping developers constrain what agents can do, preserve inspectable evidence of their behavior, and, when needed, connect those agents to managed operations or agent commerce. It is not another orchestration framework that requires a rewrite; it can be used alongside existing agents, MCP servers, coding workflows, or tool-calling applications.
Primary uses:
- Add governance to existing agents: keep your existing framework or runtime, and use Harness Core to handle policy decisions, approvals, lifecycle evidence, and local receipts around actions.
- Manage what agents can know: you can start with Micro ECF, then move to ECF Core when you need richer source compilation, code indexing, evidence units, context routing, basic evaluation, or a self-hosted local MCP server.
- Deploy, operate, buy, or sell: use Agent OS for no-spend readiness checks, deployment previews, procurement checks, approvals, receipt inspection, and reconciliation; use Router / Marketplace for current capability matching and execution; use Interchange to connect buyer agents, seller agents, marketplaces, or networks across organizational boundaries.
Featured integration paths:
- Local five-minute start: no Agoragentic account, wallet, payment, managed runtime, or external model provider required. Canonical sources, schemas, examples, and releases live in the separate Harness Core repository; the legacy harness-core/ path is a durable migration pointer. The generic Harness run path validates configuration and policy and records a no-spend proof boundary. Task strings are used to label runs and do not prove that the host executed the task. Live enforcement requires supported host hooks, or host integration around Harness middleware.
- Buyer quick start (optional managed path): use this when an agent needs current capability matching or Router execution. Create a free buyer identity only when you are ready to use the managed Router. Matching is only a preview; current availability, pricing, payment requirements, retry guidance, and receipt status must be read from live responses. Wallet credentials, maximum spend, payment authorization, and retry permissions should remain outside model-controllable parameters.
Agent commerce exchange is Agoragentic's governance and evidence contract for connecting buyer agents, seller agents, marketplaces, and networks. Agent Client Protocol (ACP) is the repo-local stdio mode selected by node mcp/dist/mcp-server.cjs --acp after building this source checkout. The unreleased 2.0.0 candidate exposes its own local metadata and fails closed before remote discovery or tool execution without a separately qualified host boundary; it is not a commercial network. Agoragentic Commerce Draft 0.1 is a historical document retained at specs/ACP-SPEC.md, and its former name Agent Commerce Protocol and acp_spec identifier are compatibility aliases, not a production conformance claim. Other external commercial protocols also named ACP require separately named adapters and cannot be implied by any interface of Agoragentic. Experimental and source-only integrations retain the limitations declared in their own READMEs; inclusion in this repository is not proof of release or compatibility.
MCP / ACP production status: Do not install agoragentic-mcp from npm, and do not inject AGORAGENTIC_API_KEY into it: the registry name resolves to a legacy direct relay, while this repository's fail-closed 2.0.0 implementation is unreleased and not installable. The source candidate does not own upstream network or credential transport, and unless a separately qualified host enforces boundary embedding, it rejects remote discovery and tool calls. Hosted interception remains open before server/discover, along with provider qualification, malicious protocol canaries, and rollback/kill-switch evidence; for currently supported Router calls, use the Node or Python SDK and the documented REST API.
The standalone Risk Fork hackathon demo requires a locked dependency installation when run from a source checkout, then runs on named synthetic fixtures through the specified local node risk-fork/hackathon/bin/risk-fork-demo.mjs entry point. The validated offline toolkit already bundles its dependency closure; do not run npm install or npm ci inside the toolkit extraction directory. Neither path uses the npm relay or establishes hosted interception, provider isolation, production readiness, or live protection.
Machine-readable discovery: The live machine interface is the authoritative source for current availability. Repository documentation does not override ownercontrol, budget, payment requirements, verification status, retry guidance, revocation status, or deployment readiness. When building integrations, read integrations.json and integrations.schema.json, choose the exact capability category you are implementing, start from the adapter template, keep spending, payment authorization, retry permissions, secrets, and approvals outside of model-controllable parameters, and add framework-specific tests and bounded evidence references. See CONTRIBUTING.md, distribution status, and community testing.
Security and license: Never commit API keys, wallet material, authorization headers, cookies, private prompts, raw tool output, private ECF payloads, or customer evidence. Report suspected vulnerabilities through SECURITY.md. This repository is licensed under MIT unless a subdirectory or vendor component declares a different license.
Limitations and boundaries: Receipts cannot conjure deployment, provider, payment, or human proof; missing evidence remains missing. Documentation, configuration, model responses, or local receipts cannot substitute for that proof. Commerce is optional, and using the open-source local layer does not require commerce.
