LocalQA: Agent for your agents

LocalQA is a Bonsai-backed local sidecar for frontier coding agents.

Your main coding agent should plan, reason, and write code. LocalQA handles the noisy local work around it: logs, browser evidence, repo context, docs, previous work memory, redaction, and context cleanup. It returns compact verified evidence so the frontier model spends less context on raw dumps.

Frontier model = planner and implementer
Bonsai = local worker model
LocalQA = sidecar layer for evidence, memory, and handoff cleanup

What we are launching

LocalQA launches with two surfaces:

Surface	Role	Who should use it first?
MCP server	Primary public interface	Anyone using an MCP-compatible coding agent
OpenCode plugin	Deeper native integration	OpenCode users who want rules, hooks, vault handles, and automatic compaction

The short version:

LocalQA is MCP-first, with an OpenCode plugin integration.

Why this exists

Coding agents are powerful, but they waste context on low-value evidence work:

Noisy input	What LocalQA does locally
Long logs and stack traces	Dedupes, ranks root-cause lines, returns compact evidence
Browser console/network/DOM dumps	Runs local Playwright QA and summarizes verified issues
Repo scans and repeated file reads	Compresses codebase evidence into source/test/risk signals
Prior work rediscovery	Writes and queries project-local memory
Raw handoff payloads	Redacts and prepares compact payloads before cloud reasoning
Tool-output bloat	In OpenCode plugin mode, can compact large tool outputs before they stay in context

MCP vs OpenCode plugin

Dimension	MCP mode	OpenCode plugin mode
Positioning	Portable default	Advanced native workflow
Main interface	MCP server over stdio	OpenCode plugin + native tools
Primary tool	local_agent_run	local_agent_run
Extra tool	Hidden in strict mode	answer_from_handle for local vault queries
Host support	Any MCP-compatible agent	OpenCode only
Tool surface	One strict public tool	Native tools plus plugin hooks
Enforcement strength	Agent chooses to call MCP	Stronger rules/hooks inside the host
Large tool-output compaction	Only when routed through LocalQA	Can happen automatically through plugin hook
Long-session memory	Returned as tool output	Rules/compaction can preserve vault/memory reminders
First-prompt privacy	No	No full first-prompt firewall
Best use case	Broad launch and easy adoption	Power users who want deeper local workflows

MCP says: the main agent can call LocalQA.

OpenCode plugin says: the agent host knows LocalQA exists and can help keep noisy tool output out of context.

How the MCP flow works

flowchart LR
  U[User gives short task or file/url pointer] --> A[Frontier coding agent]
  A --> M[LocalQA MCP server]
  M --> T[local_agent_run]
  T --> E[Local files, browser, repo, docs, memory]
  E --> B[Bonsai local worker]
  B --> C[Compact verified report]
  C --> A
  A --> P[Plan, patch, or answer]

Use MCP when you want the portable one-tool local assistant.

LOCALQA_PUBLIC_SURFACE=agent-strict npm run localqa:mcp

Tell your coding agent:

Use LocalQA as the local evidence and memory layer. Before reading large logs, browser evidence, DOM dumps, screenshots, network traces, or repeated repo context, call local_agent_run. Prefer directed mode for complex work.

How the OpenCode plugin works

flowchart LR
  U[User in OpenCode] --> O[OpenCode host]
  O --> R[LocalQA rules]
  O --> H[Plugin hooks]
  O --> LT[Native tools]
  H --> K[Compact large tool output]
  LT --> L[local_agent_run]
  LT --> Q[answer_from_handle]
  L --> B[Bonsai local worker]
  B --> V[Local vault and memory]
  V --> O

Use OpenCode mode when you want deeper integration:

Feature	What it gives you
.opencode/LOCALQA_RULES.md	Agent instructions for when to use LocalQA
.opencode/tools/local_agent_run.ts	Local sidecar agent as a native OpenCode tool
.opencode/tools/answer_from_handle.ts	Scoped follow-up questions over local evidence handles
.opencode/plugins/localqa.ts	Hooks for tool-output compaction, env injection, and compaction reminders
opencode.json	Loads the plugin and LocalQA rules from this repo

Run OpenCode from this repo root so it discovers opencode.json.

Launch proof

Latest launch check:

npm run localqa:launch-check

Benchmark	Baseline/raw	LocalQA compact	Saved	Reduction	Quality gate
Real OpenCode provider telemetry A/B	25,850 input tokens	12,048 input tokens	13,802	53.39%	One controlled Fastify task
A/B workflow proxy	25,829 tokens	6,704 tokens	19,125	74.04%	3/3 maintained
Long-horizon 300k benchmark	302,772 tokens	8,842 tokens	293,930	97.08%	9/9 quality gates, 9/9 memory writes

More detailed reports:

• Benchmarks and claims
• Provider telemetry A/B
• A/B workflow benchmark
• Long-horizon benchmark

Quick start

Install dependencies:

npm install
npx playwright install chromium

Run validation:

npm run localqa:launch-check

Start MCP strict mode:

LOCALQA_PUBLIC_SURFACE=agent-strict npm run localqa:mcp

Optional Bonsai autostart environment:

LOCALQA_DEFAULT_LLM=bonsai \
LOCALQA_BONSAI_AUTOSTART=1 \
LOCALQA_LLM_PROVIDER=openai-compatible \
LOCALQA_LLM_ENDPOINT=http://127.0.0.1:18081/v1/chat/completions \
LOCALQA_LLM_MODEL=Bonsai-8B.gguf \
LOCALQA_PUBLIC_SURFACE=agent-strict \
npm run localqa:mcp

Example tasks

Good prompt:

Use LocalQA to analyze ./logs/dev.log and return only verified auth/session errors. Do not paste raw logs into cloud context.

Good directed task shape:

{
  "controlMode": "directed",
  "task": "Investigate login QA failures and return verified evidence only.",
  "directedSteps": [
    {
      "tool": "browser_qa",
      "url": "http://localhost:3000/login",
      "task": "Collect screenshot, console, network, accessibility, and form-state evidence."
    },
    {
      "tool": "log_analysis",
      "filePath": "./logs/dev.log",
      "task": "Find auth, session, csrf, 401, 500, and stack-trace evidence."
    }
  ],
  "outputContract": {
    "maxTokens": 1600,
    "include": ["verified_findings", "evidence", "likely_files", "risks", "next_actions"],
    "exclude": ["raw_dom", "full_logs", "screenshot_metadata"]
  }
}

Bad prompt:

Here are 50,000 characters of logs pasted directly into cloud chat...

If you paste raw data into cloud chat, it already went to the cloud before MCP can help.

What LocalQA is not

Not this	Actual boundary
A frontier-model replacement	The frontier model still plans and writes code
A guarantee that pasted chat data stays local	Pasted cloud-chat text is already sent to the cloud
A patch-correctness proof system	Compression benchmarks are not patch correctness benchmarks
A formal DLP/security product	Redaction exists, but this is alpha software
Browser QA that works identically everywhere	Local browser permissions vary by machine

Public alpha docs

• Install and usage
• MCP vs OpenCode usage
• Benchmarks and claims
• Security and privacy boundary
• Launch packet

License

MIT