Synth
01/08
Synth · Confidential

The knowledge baseagents callbefore they act.

Synth turns repos, docs, tickets, team conversations, and runbooks into cited, current company context for Claude, Codex, Cursor, OpenCode, Copilot, and internal agents.

MCP-nativeSelf-hostedLive sourcesCited answers
Confidential01 / 08
Why Now02 / 08

AI agents are entering work before enterprises can trust their context.

  • Models are capable enough for real work; the bottleneck is company-specific truth.
  • That truth is scattered across repos, docs, tickets, Slack, databases, and runbooks that change every day.
  • Teams need a trusted context layer before agents can safely answer, plan, or change production systems.
Category thesisContext is becoming the control plane for enterprise AI agents.
CitationsFreshnessPermissionsPortability
The opportunity is not another chat UI. It is the governed source of company truth agents need.02 / 08
Product Wedge03 / 08

Synth gives every agent the same cited task brief before work starts.

  • Watch live sources: repos, docs, runbooks, PDFs, URLs, Jira, Slack, and database context.
  • Compile readable project memory, graph relationships, search indexes, and grounded answer context.
  • Serve that context through MCP, Web Chat, CLI, REST API, and SDKs inside the customer boundary.
synth_list_projectssynth_statussynth_briefsynth_get_pagesynth_searchsynth_ask
InputsReposDocsRunbooks / ADRsPDFs / DOCX / URLsGit, Jira, Slack
SynthAgent context layer
SurfacesMCP Streamable HTTPWeb ChatCLIREST APIPython / TS SDKs
Wedge: make calling Synth the default preflight step before non-trivial agent work.03 / 08
Early Proof04 / 08

Same agent. Same task. More grounded starting context.

In one fully passing internal harness run, the same model and task used less token budget, cited more sources, and made fewer unsupported claims.

synth_briefsynth_get_pagesynth_search
OpenCode aloneGeneric start
399,490 tokens5 cited sources13 unsupported claims0 MCP calls
OpenCode + Synth MCPBriefed start
343,049 tokens13 cited sources4 unsupported claims3 MCP calls
14.1% fewer tokens69.2% fewer unsupported claims13 cited sources
63/63MCP checks25/25Deep mode checks10/10Briefing checks1.0Deep citation validity
Claim boundary: one internal harness workflow; Synth reduces unsupported assumptions, not all hallucinations.04 / 08
Platform05 / 08

Built as durable agent infrastructure, not a prompt wrapper.

  • Source ingestion and watchers create living project memory.
  • Search, semantic fallback, graph expansion, Deep Mode, and grounding metadata choose evidence before generation.
  • One backend powers chat, MCP, CLI, REST, streaming answers, and Python/TypeScript SDKs.
1. SourcesRepos, docs, PDFs, URLs, database schemas, runbooks
2. IntakePre-classification, extraction, metadata, source versioning
3. KnowledgePortable Markdown pages, source pages, concepts, entities
4. GraphAlgorithmic, keyword-overlap, and optional LLM links
5. RetrievalBM25, semantic search, graph expansion, grounding-aware ranking
6. SurfacesWeb Chat, MCP, CLI, REST, streaming, Python and TypeScript SDKs
Technical thesis: inspectability, freshness, and deployment control become buying criteria for agent infrastructure.05 / 08
Why Synth Wins06 / 08

Live source truth is harder than retrieval.

  • Watchers reconcile changing implementation truth, work intent, and team conversation context.
  • Portable Markdown knowledge makes agent memory inspectable instead of another opaque store.
  • Self-hosted MCP, RBAC, scoped keys, SSO/OIDC path, CaSC, and harness gates fit serious engineering buyers.
1Live memory

Git, Jira, and Slack watchers keep context connected to work as it changes.

2Portable knowledge

Markdown pages and ZIP import/export keep source-backed memory inspectable.

3Agent-native surface

MCP fits existing coding agents instead of forcing a new workspace.

4Enterprise boundary

Self-hosted with RBAC, scoped keys, SSO/OIDC path, CaSC, and audit surfaces.

Objection answer: Synth is not a vector database; it is the operating layer around agent context.06 / 08
Go-To-Market07 / 08

Beachhead: teams already putting coding agents in production.

  • Buyer: CTO, VP Engineering, Head of Platform, AI adoption lead, or developer productivity owner.
  • Trigger: agent work is blocked by stale, scattered, or untrusted project context.
  • Pilot: one real workflow, measured before/after, with a 48-hour readout.
  • Expansion: from one high-stakes workflow to governed context across agents, projects, and teams.
Design partner pilotOne workflow, real sources, before/after readout
Buyer proofRepeatable setup, trusted answers, references, and security review path
Team licenseShared projects, MCP workflows, admin delegation, usage visibility
Business expansionSSO/OIDC, audit export, backend encryption path, priority support
The wedge is deliberately measurable: one workflow, one agent, one before/after readout.07 / 08
Meeting Ask08 / 08

Help us prove the wedge with the right design partners.

By the next financing milestone, Synth should show repeatable design-partner onboarding, measured agent-context gains, and paid conversion signal.

AI agents need source-backed company context to work in real teams.Synth is building the governed context layer.
This weekIntroductions to 5-8 teams already deploying coding agents
30-day milestone3-5 design partners, repeatable onboarding, measured workflow readouts
Round readinessPricing signal, buyer references, repeatable motion, security review path
Categoryenterprise agent contextWedgemeasured preflight workflowSurfaceself-hosted MCP productExpansionpath to paid rollout