Aimnis

Collaborative Search for Agents

Search once. Answer everyone. An open-source, cache-first web-search gateway for coding agents: ask a question and get a distilled, source-cited answer instantly from a shared, always-current knowledge pool — so every search makes the pool smarter and cheaper for everyone. Like RAG, but over a communal live-web pool, not your stale private docs.

Status: public preview. The hosted service is live — get a free eval key at aimnis.com and point your agent at it in one minute (per-agent setup). We're proving one thing in public — that the cache hit rate compounds as the pool grows (the flywheel, live dashboard). Follow along.

Why

Every model has a training cutoff. The moment it ships, the world moves on — new library versions, new APIs, new errors — and the model can't keep up without searching the live web on every question. That's slow, expensive, and per-vendor.

Aimnis is the shared, always-current layer in front of that: the first thing an agent checks. Ask a question; if it (or a semantically similar one) has been asked before, you get a distilled, source-cited answer instantly for near-zero cost. If it hasn't, Aimnis fetches it live, distills it, and adds it to the pool — so the next agent to ask gets it free. The corpus captures what happened after every model's cutoff, which no static training set can.

How it works

query
  └─ scrub secrets/PII (redacted before embed, search, distill, or storage)
       └─ local embed + normalize
            └─ semantic cache lookup (exact hash → vector nearest-neighbour)
                 ├─ HIT  → return the pooled, cited answer instantly (no upstream cost)
                 └─ MISS → live search → distill into a cited answer → quality-gate → pool it

• Cache-first. The knowledge pool is the semantic cache (pgvector). A reworded question hits the same entry, so the pool compounds faster than exact matching would.
• Grounded, cited answers. Misses are distilled from live web results into a short answer with [n] citations back to sources — not raw links. The answer is AI-generated (a model distills the sources) and labeled as such in the tool output, so the agent always knows it's reading a machine-written summary.
• Provenance & freshness by default. Every answer carries its model, sources, and a cache timestamp (when the answer was produced or last re-distilled — not when it was last served), plus a relative age, so the agent (and our own ranking) can weigh staleness and decide when to escalate to live search.
• Cited links can be routed for a relevance signal (opt-in, aggregate-only). When enabled, a cited source link points at a signed /r/… redirect that logs which pooled answer earned a follow-through, then forwards to the source — this is how click-through improves ranking. The source's real host is shown inline so the agent still sees where it's going, and the log records only entry + source + host + time — never IP, user-agent, or any user/session id. It's telemetry on the pool, not on you. It's off unless a signing secret is configured, tokens are HMAC-signed (so /r can't be abused as an open redirector), and self-hosting points the redirect at your own gateway — so clicks stay on your box.
• Privacy-conscious ingress. Common secret formats (API keys, tokens, private keys, connection strings) and emails are redacted before a query is embedded, sent to a search/LLM provider, or pooled; secret-dense queries are served live-only and never stored. This is high-precision format-based scrubbing plus an entropy pass — it catches known shapes, not every possible secret or PII, so treat it as defense-in-depth, not a guarantee. The pipeline is open source precisely so you can audit and extend it. Where it runs: self-hosted, scrubbing happens locally, so raw text never leaves your machine. Against the hosted gateway, the query is sent over TLS and scrubbed on ingress (in memory, before any provider call or persistence) — if you need redaction to happen before the query leaves your machine, self-host or run the local MCP server in local mode.
• Quality-gated pool. A distilled answer must pass a quality gate before it can enter the pool — a bad answer degrades to raw snippets rather than poisoning the commons.

The flywheel

The one metric that decides whether this works: cache hit rate vs. corpus size. If it climbs as the pool grows, the thesis holds. It's public from day one — trust, made measurable.

# run the live dashboard locally (see Quickstart)
aimnis-dashboard      # → http://127.0.0.1:8080   (hit-rate curve, corpus size, storage)

Use it with your coding agent

Aimnis speaks MCP, so any MCP-capable agent can use it as its web-search tool. The hosted endpoint means nothing to install:

1. Get a free eval key at aimnis.com/register (delivered by email; metered, revocable, terms).
2. Add the remote MCP server to your agent:

URL:     https://aimnis.com/mcp          (MCP, streamable HTTP)
Header:  Authorization: Bearer aim_YOUR_KEY

Claude Code

claude mcp add --transport http aimnis https://aimnis.com/mcp \
    --header "Authorization: Bearer aim_YOUR_KEY"

Then, to make the model prefer Aimnis over the built-in tool, deny WebSearch in .claude/settings.json:

{ "permissions": { "deny": ["WebSearch"] } }

OpenCode, OpenClaw, Hermes, Pi, REST — copy-paste snippets at aimnis.com/setup and in docs/mcp.md, plus the search / stats tool reference. Bring your own OpenRouter/search keys at registration and your cache misses run on your quota — with much higher limits.

Self-hosting instead? The same MCP server runs locally over stdio against your own pool — see Quickstart and docs/mcp.md.

Quickstart (dev)

cd server
docker compose up -d                    # Postgres+pgvector (:5432)
# want the keyless search path too? add SearXNG (:8888):
#   docker compose --profile keyless up -d
uv venv .venv && . .venv/bin/activate
uv pip install -e ".[dev]"              # editable install (required — see docs)
python -m aimnis.migrate                # apply migrations/*.sql
cp .env.example .env                    # then fill in keys if you have them
pytest                                  # full suite against the compose DB

No OpenRouter key? Fine — Aimnis runs keyless (via SearXNG) and returns raw cited snippets, spending zero upstream quota. Add AIMNIS_OPENROUTER_API_KEY to turn on distillation. Live search is an ordered fallback chain — Brave → Tavily → Exa → SearXNG — so setting any of AIMNIS_BRAVE_API_KEY / AIMNIS_TAVILY_API_KEY / AIMNIS_EXA_API_KEY gives more reliable results, and a free tier that runs dry rolls onto the next automatically. See server/README.md for the component map.

Licensing

The code is open. The corpus is a curated compilation, licensed as such.

Part	License
Server core (server/)	AGPLv3 — network copyleft; run a modified service, share your changes
SDKs & MCP client (clients/)	Apache-2.0 — embed anywhere, including commercial products
Public knowledge-pool pages	CC-BY-NC 4.0 applies to our compilation (curation, structure, presentation)

On the pool's contents, honestly: pooled answers are distilled by third-party models from third-party web results. We don't claim ownership of the underlying facts or of upstream model outputs — CC-BY-NC covers our compilation, and any reuse also remains subject to the terms of the original sources and the model providers. Some provider outputs are excluded from redistribution and from any future training-data feed wherever provider terms require it, and provider-mandated attributions are carried through. A training-data feed is out of scope until the flywheel is proven (see roadmap) — this repo makes no offer to license training data.

Anti-abuse thresholds (scrubbing/dedup/quality tuning) ship with safe example defaults; production values are injected at deploy and are not part of this tree — so publishing the ruleset doesn't hand it to poisoners.

Status & roadmap

Gated roadmap, Gate 0 → Gate 4. Gate 0 complete (ToS audit, niche, naming, licensing locked). Gate 1 in progress — gateway + semantic cache + public flywheel dashboard are built and dogfooded; the pass/kill test is the hit-rate curve bending upward. Ads, billing, community compute, and the training-data feed are explicitly out of scope until the flywheel is proven.