Self-hosted agent knowledge for engineering teams
A self-hosted knowledge base for AI agents.
Synth turns your repos, docs, runbooks, tickets, and decisions into MCP-accessible, cited context, so coding agents answer from company knowledge instead of guessing.
Cited contextSelf-hosted · MCP-native · cited answers · engineering-first
Benchmark previewSame agent. Same task. One run had Synth context.
With Synth MCP: 9 cited source paths instead of 1, real MCP context tools used, and a passed trust gate.
MeasureNo SynthWith Synth
Company context
Agent guesses
Agent retrieves cited context
Cited source paths
1
9cited
MCP context tools
None
Used
Trust gate
Failed
Passed
Time to useful plan
33.9s
20.5s39.6% faster
LLM tokens
218.7k
135.6k38.0% fewer
Built for teams already using Cursor, Claude, Copilot, Windsurf, OpenCode, or internal agents.
The problem
Your agents do not know how your company works.
Teams are already using AI agents, but the critical context is scattered across code, docs, tickets, incidents, and people. That makes agents guess.
What is Synth?
A self-hosted knowledge base for AI agents.
Synth is a self-hosted knowledge base for AI agents. It serves cited company context through MCP before agents answer or act.
Connect the sources that explain how your company really builds and operates.
Agents retrieve the right brief from Synth inside the workflows they already use.
The answer includes receipts, not just confident language.
Private knowledge layerOne cited context layer for every agent.
Buyer workflows
Prove value on real engineering work.
Start with one high-value workflow and measure whether agents become more grounded, cited, and ready to act on your own material.
Agents retrieve architecture context, gotchas, ownership, and validation steps before editing.
Agents find runbooks, past incidents, known failure modes, and recovery guidance.
New engineers get cited read-first paths, team conventions, and subsystem context.
Leaders see source coverage, gaps, trust gates, and rollout recommendations before scaling.
Product surface
The answer has to show its work.
The useful surface is concrete: what context was retrieved, which sources backed it, what risks were found, and whether the answer is safe to use.
Synth Control Planeteam-context / readiness-pack
Get the team playbook first.
Attach source paths to guidance.
Flag weak grounding before rollout.
Cited paths9
Token spend135,583
Unsupported claims1
StatusWARN visible
The benchmark task
The proof uses one real planning task.
After the product is clear, the benchmark shows the difference. We asked OpenCode to plan a developer task. The only difference: one run could call Synth MCP first and retrieve trusted project memory.
Actual promptThe exact task given to OpenCode
Tell me step by step how to create a new Synth harness to improve a component. Include the folder layout, files to write, metrics to track, validation gates, and mistakes to avoid.
Same-agent proof
Same agent. Same task. One run had Synth context.
OpenCode alone gave useful generic advice. OpenCode with Synth MCP retrieved project memory, cited more real paths, used context tools, and passed the answer trust gate.
Run AOpenCode alone
Output
Good generic harness advice.
Grounding
1 cited source.
Trust gate
Failed.
Cited paths1
Unsupported claims3
MCP toolsNo
Run BOpenCode + Synth MCP
Output
Cited team workflow.
Grounding
9 cited source paths.
MCP tools
synth_prepare_task and synth_get_page.
Cited paths9
Unsupported claims1
Trust gatePassed
Measured comparison
What we tested and what improved.
Both runs received the same prompt: plan the folder layout, files, metrics, validation gates, and mistakes to avoid for a new Synth harness. The harness then checked whether the answer was grounded, cited, project-aware, free of hallucinated file paths, actually used Synth MCP, and safe to count on.
A developer should be able to follow the answer without another prompt.
Same OpenCode agent, same repo, same model, same task. One run retrieved context first.
With Synth MCP, the agent cited 9 source paths instead of 1, used real MCP context tools, reduced weak claims, and passed the trust gate.
Highest trusted resultWARN
- Total LLM tokens
- 218,677 -> 135,583
- Token savings
- 38.0% fewer
- Time to useful plan
- 33.930s -> 20.481s
- Useful plan speed
- 39.6% faster
- Cited source paths
- 1 -> 9
- Unsupported claims
- 3 -> 1
- Real MCP context tools
- synth_prepare_task + synth_get_page
- Answer trust gate
- failed -> passed
Highest fully clean PASS resultPASS
- Total LLM tokens
- 399,490 -> 343,049
- Token savings
- 14.1% fewer
- Unsupported claims
- 13 -> 4
- Public claim
- Safest number
- Run status
- Fully passed
Latest two additional runsTRACKED
- Run 1
- PASS · 3.0% fewer tokens
- Run 2
- WARN · -8.3% token savings
- Efficiency claim
- Up to 38% fewer tokens and 39.6% faster to useful plan in a trusted same-agent comparison.
- Grounding claim
- More grounded answers: 9 cited source paths, real MCP context tools, and a passed trust gate.
- Conservative claim
- 14% fewer tokens in a fully passing same-agent comparison.
The honest caveat
The warning is part of the proof.
The strongest run passed the answer trust gate but was marked WARN because the harness caught small pre-MCP narration. This is good: Synth measures the process honestly instead of hiding imperfections.
Quality rule
Only count a win when the answer passes the trust gate: grounded, cited, project-aware, no hallucinated file paths, and actual Synth MCP tool usage.
Governance
Built for teams that need control, not another black box.
AI adoption gets messy when agents can act without inspectable context. Synth keeps private knowledge cited, inspectable, and measurable.
Private engineering context stays under your control.
Source paths, validation steps, and risks stay attached.
One context layer can serve multiple coding tools.
Grounding, tool usage, and process quality can be scored.
SSO, audit retention, admin controls, and support fit naturally as rollout grows.
48-hour pilot
Deploy agent-ready company knowledge in 48 hours.
In two days, Synth can deploy privately, ingest one high-value knowledge set, connect one agent workflow, and prove cited answers on your own material.
Self-hosted or controlled pilot environment.
Repos, docs, runbooks, tickets, or decisions.
One agent workflow calls Synth before answering.
Proof on your own engineering material.
Knowledge gaps, trust signals, and next-team plan.
FAQ
Questions serious buyers ask.
No. Synth gives those agents private company context through MCP.
Docs store knowledge. Synth packages cited, task-specific context for agents before they answer or act.
Yes. Synth is designed for private deployment or a controlled pilot environment.
No. The primary value is grounded, inspectable answers. Efficiency is measured when it appears.
One team, one workflow, and 1 to 3 high-value source sets such as repos, docs, runbooks, or tickets.
Next step
Give your agents company context before they act.
Start with one team, one repo set, and one workflow. In 48 hours, prove whether Synth can improve grounding, citations, and agent readiness on your own material.