Question 1

What exactly is CodeOrch?

Accepted Answer

CodeOrch is a persistent runtime layer that sits between your team and your AI coding agents. It gives Claude Code, Cursor, Aider, and any MCP-compatible agent a shared memory, so every decision, every file touched, every architectural choice compounds into the next task instead of evaporating at session end.

Question 2

How does the 70%+ token reduction actually work?

Accepted Answer

When an agent claims a task, the orchestrator fires three parallel queries: the code graph for relationships and file dependencies, the vector store for semantic recall of similar past work, and the task database for explicit architectural decisions. These collapse into a 2–6k token surgical briefing. The agent reads the briefing, not 28k tokens of raw codebase. Pre-checks skip files where the grep already confirms nothing changed.

Question 3

How is this different from just using MCP directly?

Accepted Answer

MCP is a protocol, it connects agents to tools. CodeOrch is an orchestrator, it manages what context those tools return and in what order. Raw MCP gives your agent a file-reader and a grep tool. CodeOrch hands it a ranked, deduplicated, relationship-aware briefing that took 2 seconds to synthesize from three databases. One is plumbing; the other is architecture.

Question 4

How long does it take to set up?

Accepted Answer

About 30 seconds for cloud. Add two lines to your .mcp.json, restart your agent, and the runtime is live. For self-hosted, docker compose up starts the backend, database, graph, and vector store. Your first orchestrated task typically runs within 5 minutes of signup.

Question 5

Which agents are supported?

Accepted Answer

Any agent that speaks MCP. Claude Code, Cursor, Aider, and Opencode are first-class, with tested .mcp.json configs in our docs. Custom agents that implement the MCP spec work too. If your agent can call tools over stdio or SSE, CodeOrch works with it.

Question 6

Can multiple agents work on the same codebase at once?

Accepted Answer

Yes, that's a core use case. When Agent A claims auth/*, CodeOrch records the file lock and injects it into Agent B's briefing automatically. Discoveries flow both ways: if Agent A finds a JWT signing bug, Agent B sees it in its next claim. Merge conflicts caused by agent overlap drop to near zero.

Question 7

Is my code or my tokens ever stored on your servers?

Accepted Answer

We never store your source code. Only indexed relationships (code graph), embeddings (vector store), and decision records (task database) are stored, none of which reconstruct raw code. GitHub PATs are Fernet-encrypted at rest with KMS-versioned keys. WebSocket ticket exchange means no JWT ever touches browser history. Every mutation has an append-only audit log.

Question 8

What counts as an orchestration?

Accepted Answer

One orchestration = one full begin() → finish() task cycle. Browsing the dashboard, searching the graph, reading decisions, none of that counts. Free tier gives you 100 orchestrations/month which is roughly 3–4 active tasks per day. Pro and Business tiers scale to 1,000 and 10,000 respectively.

Agents haveamnesia.We fix that.

From idea to PR, fully wired.

It lives where your agents live.

Same task. Different runtime.

Built to compound.

Questions,
answered.

Give your agents a memory.