Synapse

A multi-agent MCP orchestrator for Solana DeFi. Send a task — Synapse routes it to the right combination of specialized Claude agents, runs them in sequence, and returns a synthesized action plan.

How it works

Most DeFi analysis requires at least two distinct skill sets: reading raw on-chain state (pool TVL, bin arrays, prices) and interpreting what that state means in market context (trend, funding sentiment, cross-exchange divergence). These are different jobs that benefit from different prompts, different tools, and different reasoning styles.

Synapse splits them:

• defi-agent — on-chain data specialist. Fetches pool states, scores volume authenticity, pulls token prices. Reports raw numbers.
• market-agent — market analyst. Reads funding rates, classifies price trends, flags divergences between spot and perpetuals.
• executor — synthesizer. Receives the other agents' output as context and returns a ranked action plan with confidence scores.

A Claude-powered router decides which agents a given task actually needs — no hardcoded rules, no over-dispatching.

Routing

The router is itself a Claude call. Given your task, it selects the minimum set of agents needed:

"Get current SOL/USDC pool state"
  → defi-agent only

"Is the market trending bullish?"
  → market-agent only

"Analyze top pools and recommend actions"
  → defi-agent → market-agent → executor

This avoids burning tokens on agents that aren't relevant to the task.

Dashboard

A live status dashboard runs at http://localhost:3000 while the orchestrator is active. It shows the last cycle's agent invocations, tool call counts, timings, and executor output. Auto-refreshes every 10 seconds.

Quickstart

git clone https://github.com/DeltaLogicLabs/synapse
cd synapse
bun install
cp .env.example .env    # add your API keys
bun run dev

Dashboard: http://localhost:3000

Configuration

Adding an agent

1. Create agents/your-agent.ts extending BaseAgent
2. Implement definition, getTools(), executeTool()
3. Register in agents/base.ts → AGENT_REGISTRY and createAgent()
4. Add to the router's system prompt in core/router.ts

The router will automatically start considering your new agent for task dispatch.

Technical Spec

Forced Dispatch — tool_choice: "any"

The router uses tool_choice: { type: "any" } on its Claude call, which forces the model to call dispatch_agents rather than returning a plain text response. Without this, ambiguous tasks ("what's happening in DeFi today?") may elicit a prose answer instead of a routing decision, causing the fallback to over-dispatch all three agents and waste tokens.

tool_choice: { type: "any" }
// router MUST call dispatch_agents — no text-only escape hatch

Context Accumulation Pattern

Each agent in the sequence receives the prior agents' outputs as context, not just the original prompt. The executor's input is:

sharedContext          ← ContextStore.toPromptContext()
  + defi-agent output  ← raw on-chain numbers
  + market-agent output ← trend/funding interpretation
  + original task prompt

This lets the executor synthesize without needing its own tool calls to fetch on-chain state — it only needs rank_actions to emit its verdict.

Context Size Guard

ContextStore.toPromptContext() truncates at 4,000 characters. In long-running sessions, recentDecisions accumulates 50 entries — even at 100 chars each that's 5,000 chars of prior decisions injected into every agent's input. The guard prevents token bleed from old decisions crowding out the actual task.

Token Budget Tracking

Token usage is accumulated across all turns in each agent's multi-turn loop and returned in AgentResult.tokenUsage. The orchestrator sums this at cycle end:

const cycleInputTokens = cycle.results.reduce((sum, r) => sum + (r.tokenUsage?.inputTokens ?? 0), 0);

This lets operators track actual API spend per cycle without external billing aggregation.

Agent Timeout

AGENT_TIMEOUT_MS (default 60s) bounds each agent's execution. The check runs at the top of the while (true) agent loop — before each API call — so a hung previous turn doesn't trap the orchestrator waiting for a response that will arrive after the cycle window closes.

Confidence Gate on Routing

If the router's own confidence field falls below CONFIDENCE_THRESHOLD, the executor is dropped from the dispatch list. This prevents action plans from being generated when the task scope is ambiguous — the executor's rank_actions output is only reliable when defi-agent and market-agent have been given a well-scoped task.

Stack

• Runtime: Bun 1.2
• Agents: Claude Agent SDK — individual stop_reason === "tool_use" loops per agent
• Routing: Claude with tool_choice: "any" — forced dispatch decision
• Dashboard: Bun native HTTP server with token usage per agent

License

MIT