THE AGENTIC LAYER
Three kinds of intelligence.
One set of principles.
Saroir is built with agents from day one, but on terms that respect contributors and the record. A platform agent keeps the knowledge graph alive. A project agent holds each project's memory. And any external AI can connect through an open protocol, under your control.
01 · SAROIRBOT
The platform's ever-present intelligence, continuously connecting knowledge across every contribution.
Saroirbot reads every public project, every accepted contribution, every cohort report across Saroir. Its job is to connect work to work: to catch when a cohort is repeating an approach that has already been tried, to surface prior art when a new ticket is created, and to show where knowledge gaps persist across the platform.
Cohort briefing
When a cohort forms, Saroirbot writes a short brief summarising what the project has tried before. Prior approaches, ruled-out paths, open questions. The cohort starts from something, not from scratch.
Cross-domain bridge
When a new ticket is posted, Saroirbot surfaces related work from other domains. A question in applied ML that was answered two years ago in operations research no longer gets solved twice.
Knowledge gap signal
Saroirbot identifies unsolved problems that persist across the platform. Projects cluster around them. Moderators flag them for Hot Project designation.
Saroirbot does not generate contributions. It points, summarises, and cites. The work is still human work.
Saroirbot operates under platform guardrails. It reads only public data. It cannot resolve a DID to a human. It cannot access identity database. Every action it takes is logged.
02 · PROJECTBOT
One agent per project. Institutional memory that outlasts the people.
Projects on Saroir may last months or in some cases, years. Cohorts come and go. Contributors hand off. Projectbot is the continuity layer: one agent per project, created automatically at project approval, holding everything the project has learned across its full lifespan including periods of hibernation.
Continuity across cohorts
A new cohort joining a three-year-old project does not need to read every cohort report. Projectbot summarises where the work is, what has been tried, what the current cohort should know.
Citation and challenge
When a new cohort proposes an approach, Projectbot checks it against prior work. If the same path has been tried before, it surfaces the earlier cohort report. If the approach is genuinely new, it says so.
Timeline narrative
Projectbot can write human-readable summaries of the project timeline for funders, new contributors, or the public project page. Always labelled as agent-generated. Contributors edit before publishing.
Each Projectbot is governed by the project's moderator, not by any single contributor. It cannot submit work. It cannot vote. It remembers.
Projectbot has its own identity on the platform, created at project approval and tied to the project for its entire lifespan. It reads all contributions, cohort reports, and documents within the project, but nothing outside it. Its guardrails are set collectively by the founding contributor and the first cohort leads. All activity is on the audit log.
03 · MCP SERVER
Connect your own AI.
Keep everything portable.
Saroir exposes its data through Model Context Protocol, the open standard for connecting AI assistants to external systems. Any MCP-compatible client can query your contributions, your credentials, your cohorts, and the public knowledge graph, under permissions you control.
Work with Saroir from your own tools
Claude Desktop, Cursor, VS Code, ChatGPT with custom connectors, and any other MCP client can read and write to Saroir on your behalf. Draft a contribution in your editor. Query your record from your assistant. Submit when ready.
Scoped access you control
Every MCP connection uses a token scoped to what you explicitly permit. Read-only access to your credentials. Write access to contribution drafts. Access to a specific project's workspace. The scopes are public and revocable. Your DID holds the keys.
No lock-in, no vendor dependency
MCP is an open protocol. You are not tied to a specific AI vendor or IDE. Switch clients any time. Your Saroir data stays on Saroir. Your conversations stay in your tool. The connection is the only thing that moves.
Your credentials are portable. Your data is portable. Your tools should be too.
MCP access is available to all contributors from Mark 1. Every request through the MCP server is authenticated, scoped, and logged. You can revoke any token at any time. No third party sees your Saroir data except the AI client you explicitly connect.
Companies with founding partner status get programmatic MCP access to their sponsored projects, with scopes set by the project founder.
How the three agents work together
The three agents operate at different scopes. MCP sees what you permit. Projectbot sees its own project. Saroirbot sees the whole public graph. They interact through a structured message layer, not through direct data access.
Every agent action is logged. No agent writes to the contribution record directly. No agent resolves a DID to a human identity. Agents interact with the message layer, not with identity database.
PRINCIPLES
The same rules apply to every agent on Saroir.
Human judgment is the decision. Every agent points, summarises, and suggests. Contributors decide what gets submitted, merged, and attributed.
Agents see only what their scope allows. Saroirbot sees public work. Projectbot sees its own project. Your MCP client sees what you permit. No agent crosses into identity database.
Every agent action is logged. Agents have their own DIDs. Their actions appear on the audit log. Contributors can review what was done on their behalf.
No agent can un-do a contribution. Credentials are permanent once accepted. Agents assist in drafting and reviewing. They do not edit history.
FAQ
“What is the difference between Saroirbot, Projectbot, and MCP?”
Saroirbot and Projectbot are Saroir's own agents, built into the platform and governed by moderators. MCP is how you bring your own AI assistant to work with Saroir. The first two know the platform. The third is yours.
“Do I have to use an agent?”
No. Every Saroir feature is usable by humans alone. Agents are opt-in assistance, never a requirement.
“What can my MCP client see?”
Only what your token permits. By default, a new token has no scopes. You add what you want: your own credentials, your active projects, your cohort workspaces. Public data is always readable.
“Can my MCP client submit a contribution on my behalf?”
It can prepare a draft, but final submission requires your explicit confirmation through the Saroir interface. This is structural, not a toggle.
“Do you support other AI vendors?”
Yes. MCP is an open protocol supported by Claude, ChatGPT, Cursor, Continue, and a growing list of clients. Saroir's MCP server is vendor-neutral. Any compliant client works.
Built on open-source infrastructure and a strong belief that knowledge should be accessible to everyone.