fyrr

mcp
Security Audit
Fail
Health Pass
  • License — License: Apache-2.0
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 10 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in .github/workflows/ci.yml
  • rm -rf — Recursive force deletion command in .github/workflows/release.yml
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

The serverless, permission-aware context layer for AI agents. No database. Just a bucket.

README.md

fyrr: serverless context layer for AI agents. No database. Just a bucket. The logo over a dithered sea chart of knowledge islands named after the apps it indexes.

The serverless, permission-aware context layer for AI agents. No database. Just a bucket.

Website · Docs · Use cases · Compare · Changelog

License: Apache-2.0 Rust 1.97+

Pre-1.0. Cloud connectors are tested against synthetic APIs only.

Why fyrr

  • Find: works in the AI you already use. Plug fyrr into Claude, Cursor or
    any MCP app. Ask there, and it searches your docs, chats, tickets and code
    over stdio or Streamable HTTP (MCP tools).
  • Protect: only what each person may read. fyrr syncs each document's
    access rules with its content and checks them inside the search, before
    anything is ranked. A forbidden document looks exactly like one that does
    not exist, and missing permission data refuses rather than guesses
    (how, security).
  • Scale: as many readers as you need. Readers only read from the bucket
    and never coordinate, so you add more as you grow
    (topology).
  • Own: your data stays yours. It runs in your own cloud. Text leaves it
    only for the embedding model you choose.

Connectors for local files, Google Drive, Notion, GitHub, Slack and Jira; a
Python SDK lets you write your own.

16 milliseconds. Among the best results.

Fast enough for an agent to search at every step: a 16 ms median search, 25 ms
at p95. Accurate enough to trust what comes back: a 90+ score on
EnterpriseRAG-Bench, among the best of 43 systems tested. The
open comparison measures quality, latency, ingest
and running cost against self-hosted alternatives, with its full protocol.

No database. Just a bucket.

Everything fyrr knows lives as files in one S3 bucket, or in a private local
directory. Small teams run nothing around the clock; a whole company adds one
small service. The reader also runs on AWS Lambda
(experimental): no always-on reader server, pay per request.

Sources sync through the writer into S3 or local disk; the reader checks permissions before returning MCP results.

From one laptop to the whole company.

Way What runs
Just you Your files, notes and repos, searched by your coding agent.
A team One shared index, no server. Each teammate's agent searches it directly.
The company One index and one server, behind your company login.
Training & evals A fixed index that changes only when you update it. Search it in parallel.

Every way: no database. See use cases and
deployment.

Quick start

Follow the quickstart to search a Markdown folder from
Claude Code using a private local directory. It needs Linux x86-64 and the
tools git, curl, unzip, openssl, Python 3.11+, rustup, uv, a C compiler, jq,
binutils, network access and about 10 GiB free; no accounts
or embedding keys are needed. The guide also covers adding Slack and trying
Ada/Grace permissions. For shared deployments, see
deployment.

After setup in the quickstart:

fyrr writer namespace create
fyrr writer sync
fyrr writer commit
claude mcp add fyrr --scope user -e FYRR_IDENTITY_KEY="$FYRR_IDENTITY_KEY" -- "$PWD/target/debug/fyrr" --config "$FYRR_CONFIG" reader serve

Ask Claude:

Use fyrr search and fetch to answer: How do I restore a backup? Quote the handbook.

The example's top hit is notes/backup.md, with the answer:
"To restore a backup, run the restore job and verify the checksum."

Limits

  • You need a stable release. fyrr is pre-1.0: storage formats and the
    MCP schema may break before 1.0; data is then republished, not migrated.
    Releases (binaries, images, the SDK wheel) will be published from v* tags.
  • You need proven cloud connectors. Google Drive, Notion, GitHub, Slack
    and Jira have been tested only against synthetic APIs, not real workspaces.
  • Your documents are not text. There is no PDF, Office or image
    extraction.
  • You cannot send text to an embeddings service. You provide an
    OpenAI-compatible /embeddings endpoint, and document and query text is
    sent to it.
  • You do not run Linux. Linux (x86_64, aarch64) is the only supported
    platform, for builds, binaries and images.
  • You need phrase queries. The full-text index stores no term positions.
  • You want TLS built in. The HTTP listener speaks plain HTTP; outside
    loopback it needs a TLS-terminating proxy.

Documentation

Page Content
Quickstart Build, configure and search a local folder
How fyrr works Terms, publication, authorization and retrieval
Security model Trust boundaries, permissions and secrets
HTTP login and users Authenticate and enroll HTTP users
Deployment Production setup, images and scheduling
AWS Lambda Experimental reader deployment
Operations Monitor, restore, retire and upgrade
Configuration Storage, embeddings, sources and secrets
Identity Users, directories, keys and OIDC
Tuning Namespace and retrieval settings
Reader and writer settings Serving, retention and maintenance settings
CLI Commands, flags and exit statuses
MCP tools Tool arguments, results and transport
Connectors Setup and limits per source
Writing a provider Connector SDK contract and example

Contributing

Bug reports, connector requests and pull requests are welcome; for larger
changes, open an issue
first. CONTRIBUTING.md covers the build, scripts/ci and
the contract rules. Report vulnerabilities privately as described in
SECURITY.md, and follow the
code of conduct. Notable changes are listed in the
changelog.

License

fyrr is licensed under the Apache License 2.0; see NOTICE.

Reviews (0)

No results found