Self-hosted deterministic knowledge for AI agents

Give every AI agent the same trusted Project knowledge.

Synth prepares your repositories, documentation, runbooks, and decisions into durable Project knowledge. Agents search trusted knowledge, open exact pages, and do their own reasoning through MCP.

Same model. Same task.Up to 53% fewer research tokens · Up to 71% fewer tool calls

Plan a Pilot See how it works

How Synth works

Agents can generate. They still need your real knowledge.

AI agents often begin work without your architecture decisions, runbooks, standards, prior incidents, or current Project knowledge. Synth creates a deterministic path from trusted knowledge to agent action.

01Prepare trusted knowledge

Turn repositories, documentation, runbooks, and other sources into durable Project pages.

02Publish trusted knowledge

Successful changes make a consistent Project view available to every connected agent.

03Search and open exact pages

Agents receive stable ranked excerpts and open exact pages from the active Project.

04Let the agent think

The connected agent reasons, plans, codes, and cites. Synth does not generate its answer.

Synth retrieves trusted knowledge. The connected agent produces the answer.

Synth in action

Search the Project. Read the exact page.

This example use case makes the boundary visible: Synth finds the right Project page, then the external agent writes the answer.

You: How do I run the app locally? Use only Synth MCP.

→synth_list_projects()→synth_search(project: "web-app", query: "local setup")→synth_get_page(page: "docs/getting-started/local-development.md")Agent:1. Run npm install
2. Copy .env.example to .env
3. Run docker compose up -d
4. Run npm run dev · Open localhost:3000

Source: docs/getting-started/local-development.md

Measured on the same task

Less research. Fewer tool calls.

Compared with repository-only research, Synth MCP cut token use by up to 53% and tool calls by up to 71% on the same task.

Up to53%Fewer research tokens

Up to71%Fewer tool calls

OpenCode agent comparison · same model · same task · Synth MCP versus repository-only research

One contract, every surface

Read the same Project knowledge wherever agents work.

MCP, Explore, CLI, REST, and SDK expose the same Project-scoped retrieval contract without adding a hidden answer layer.

MCPExploreCLIRESTSDK

Read-only MCP contract

Exactly four tools.

Discover Projects, retrieve ranked sources, read exact pages, and inspect readiness.

01synth_list_projects02synth_search03synth_get_page04synth_status

Pilot

Prove it on one Project and one real workflow.

Connect a trusted source set, give one agent Project-scoped access, and compare how it performs with and without Synth knowledge.

Plan a Pilot See the workflow

FAQ

Clear boundaries build trust.

01Does Synth replace Claude, Cursor, Codex, or Copilot?

No. Synth gives those agents trusted Project knowledge. They remain responsible for reasoning and acting.

02Does Synth generate answers?

No. Synth retrieves ranked sources and exact pages. The connected agent generates the answer and cites what it used.

03Does retrieval call an LLM?

No. Canonical search and exact page reads are provider-free.

04Why not use ordinary internal search?

Synth adds Project authorization, deterministic ranking, stable source identity, exact page reads, and agent-native access through MCP.

05Can we self-host it?

Yes. Self-hosted control and Project isolation are central to the product.

06Is the knowledge portable?

Yes. OKF-compatible import and export and human-readable knowledge support portability and inspection.

Next step

Give every agent the same trusted Project knowledge.

Start with one Project, one source set, and one real workflow.

Plan a Pilot See the MCP contract