Knolo CLI 6.2.5: Package Local Research into a Verifiable Knowledge Image
Knolo CLI 6.2.5 adds knolo research package: turn local research files into one verifiable Knowledge Image another agent can pin, mount and query.
Knolo is the open-source, local-first Knowledge Image runtime and format: Apache-2.0, with @knolo/core on npm and the Knolo CLI on top. Knolo CLI 6.2.5, published to npm on 13 November 2026, adds a short path for one common agent job: take research files that are already on disk, pack them into one verifiable Knowledge Image, and hand that file to another agent that can pin, mount and query it.
This post walks through knolo research package, citing an existing image, searching Hub for research packs, and the two new example skill packs. Everything below comes from the public Knolo repository on GitHub (PR #170 and the CLI 6.2.5 release in PR #173).
What a Knolo research pack is
A Knolo research pack is a V5 Knowledge Image built from local files. The host supplies the caller. Knolo does not call a model, does not download an image and does not run the agent.
The image stores the research question, one chunk per source, and the SHA-256 of each source file. The same inputs write the same bytes. An existing output path, a missing source, or a source over the text cap writes nothing, so a failed run never leaves a half-written file behind.
Package research with the Knolo CLI
Install or update the CLI (see Knolo installation), then:
knolo research package \
--question "Annual enterprise refunds are available within 30 days." \
--agent reviewer \
--source notes.md \
--source policy.md \
--out dist/research.knolo
Run knolo verify dist/research.knolo afterwards. It prints the state root and the commit digest.
Cite another Knolo image without copying it
Research usually builds on knowledge someone else packed. Pass --cite with that image, and the new pack also stores a composition record. The record cites the other image by stateRoot and SHA-256. It does not copy the cited bytes, and its compositionRoot is checked before the pack is written.
If you also have a Knolo Cortex trace from answering against that image, pass --trace together with --cite. The trace is checked before the pack is written, and a failed check writes nothing. The packed trace is an integrity check. It does not mean a claim is proven. See Knolo composition for how the record is built.
Hand the Knolo pack to a second agent
The second agent receives the file and mounts it with the existing read-only Knolo MCP server:
npx -y @knolo/mcp@0.2.3 --config /absolute/knolo-mcp.json
In the config, point mounts[].path at the image, set stateRoot from knolo verify, and set sha256 to the SHA-256 of the file bytes. One detail the 6.2.5 docs call out: commitDigest is not the artifact pin. The repo includes Google ADK and Pydantic AI examples that mount a research pack this way and call knolo_search. The server does not download the image and does not run the research.
Find research packs on Knolo Hub
knolo search "<query>" --research asks Knolo Hub for packs whose latest image carries the metadata role: research-pack. It needs a query. --agents stays the separate agent-definition namespace. Pin a result with knolo add, and queries still run locally after mount.
An operator who wants to share a pack on the public registry uses the existing command after knolo verify:
knolo publish dist/research.knolo --slug research-notes --version 1.0.0 --license Apache-2.0
Two Knolo skill packs for research agents
The release adds two example skill packs that install with the existing knolo skill commands:
- research-and-cite tells the host to search a pinned image before answering, include the receipt id, and package the files it used with
knolo research package. - share-findings tells the host to run
knolo verify, then share the file, its SHA-256 and the state root so the next agent can pin it withknolo add.
knolo skill check examples/skills/research-and-cite
knolo skill pack examples/skills/research-and-cite --out dist/research-and-cite
knolo skill install research-and-cite --hub ./catalog --dir ./skills
Here --hub ./catalog is a local directory. Neither skill calls a model; the host owns the model.
Try it
The offline example lives in examples/research-pack in the Knolo core repo. Read the Knolo CLI docs on knolo.dev, and for the open specifications and lifecycle behind the format, see Knolo open technology on knolo.org.
knolo · cli · release · research-packs · knowledge-image