Any Agent, Real Work: MetaSpark MCP+API Ships
Agents stop being demos when they have a harness. MetaSpark's MCP+API lets Claude, custom models, or your own agent integrations execute real work across your entire tool stack without waiting for roadmap dependency.
The Harness, Not the Agent
Every AI productivity tool in the market is actually a chatbot disguised as infrastructure. You feed it context, it generates text, you copy the output somewhere else. The real work still happens in Linear, GitHub, Slack, Notion, Gmail, the systems that are already open on your team's screens. The gap between "generated output" and "done work" is where all the friction lives.
MetaSpark v2 flips this architecture. The platform is the AgentOS, the runtime layer that sits between any agent and your actual systems of work. You bring the agent. MetaSpark gives it the ability to read everywhere, plan multi-step actions, execute, write back to source, recover cleanly when something fails, and log every move so your team stays in the loop.
That means you're no longer hostage to which model any single vendor decided to embed. If Claude's reasoning is superior for your use case, use Claude. If you've built a custom LangGraph planner optimized for your deal flow, plug it in. If you want to layer multiple agents, one for code review, one for customer ops, one for vendor relationships, they all run under the same harness, read the same signals, write back to the same source systems, and show up in the same audit log.
What Shipped
MetaSpark MCP + API is the integration layer for any external agent.
Via the Model Context Protocol, Claude (or any agent framework compatible with MCP) can connect to MetaSpark and call tools that MetaSpark exposes: read the current ranked task list, inspect dependencies, query the connection graph across Linear / GitHub / Slack / Notion / Gmail / Calendar, draft a reply to an email thread, file a ticket, writeback a status, tag a human for approval. The same toolkit our in-product agents have.
Via the public REST API, you can invoke agents programmatically. POST a task, specify which agent should run against it, get back the result with full audit trails. Your custom orchestration can wrap MetaSpark's runtime instead of building its own.
Both paths give you the same guarantees: if an agent hits a system it doesn't have a connector for (your internal compliance portal, your partner's custom REST API, a legacy ERP), it can author the integration in under a minute. Read the docs or API schema, infer the resource model, draft the OAuth flow and webhook listener if needed, verify against your real data, ship it live. No waiting for our product roadmap. No integration limbo.
Why This Matters
The collapse of monolithic AI is already here. Teams don't want "one AI vendor that does everything." They want their model choice, their orchestration choice, their cost structure, their privacy guarantees. They want to layer specialized agents, a code reviewer, a customer-ops dispatcher, a proposal writer, instead of forcing everything through a single general-purpose bot.
What they haven't had until now is the infrastructure to actually route real work to those agents. Every "AI agent" product is really a demo runner: call the API, get a generated output, figure out the integration yourself. That's fine for drafting an email. It doesn't work when your agent needs to read a 40-ticket sprint, file blockers, bump priority on dependent items, and writeback status back to source without a human watching every move.
MetaSpark's MCP + API changes that constraint. The harness exists. Any agent can use it. You own the choice of model, orchestration, and specialization. Your agent gets real execution against real systems.
How Teams Use It
A technical founder who's built a custom LangGraph agent for their specific deal flow can now plug it directly into MetaSpark instead of building its own integrations layer. Her agent reads the ranker's daily output, identifies which prospects have the highest probability of close this quarter, drafts the follow-up sequence, files the tasks, and writes back to the CRM. Same with an engineering lead who wants a specialized code-review agent that reads commit diffs against architectural guidelines before PR comment.
A larger team can deploy multiple agents: Claude for reasoning-heavy tasks, a fine-tuned model for classification, a custom rule-based planner for deterministic workflows. They all talk to MetaSpark, all read the same dependency graph, all see each other's actions in the audit log. No data silos. No integration duplication.
Getting Started
MCP support is live now. If you're running Claude desktop or using Claude in your terminal via MCP client, you can configure the MetaSpark MCP endpoint and start routing tasks through Claude directly into your MetaSpark workspace.
The public API is available for early access. POST to /agents/{agent_id}/execute with a task object, get back structured results with full audit trails. Documentation is live in our developer portal.
If your agent needs to reach a system we haven't built a connector for yet, there's no roadmap to wait on. Your agent can describe the system and ship the connector itself. Median time: 47 seconds. Slowest we've seen on a hostile internal portal: just over two minutes.
This is what agents look like when they have a harness beneath them instead of just a chat box bolted onto every tool. Read the docs, try it, and build something that actually finishes work instead of just talking about it.