LoginSign Up
← Back to all sparks
J

Jan

AI-ASSISTANTS
Velocity5.0

Open-source offline AI assistant that runs language models locally on your device.

jan.ai ↗

Jan is quietly wiring subagents into the same tool pipeline its main agent uses.

local-aisubagentsnative-toolsllama.cpp-defaultsdesktop-appstabilization
◆Current state
Jan's tagged releases in this window are small: a persisted chain-of-thought duration, a CSP fix that unblocks video uploads, and a change to llama.cpp defaults that turns auto-fit off and pins context length to 8192. Cadence is slow — four tags spanning May to July. The most recent tag is not a release at all but a development checkpoint.
◆Where it's heading
That checkpoint is the informative one: subagents now reuse the main native tool pipeline rather than a separate path, alongside code-UI work. Jan is consolidating on one tool-calling surface for both the primary agent and its subagents, which is the precondition for multi-agent workflows inside a local desktop app. The shipped releases meanwhile read as stabilization of the chat surface — durable metadata, predictable inference defaults.
◆Prediction
The subagent and code-UI work visible in the checkpoint tags should surface in the next minor release; on this cadence, expect more 0.8 stabilization patches before it does.

◆Recent moves

  1. 26d ago

    Dev checkpoint: subagents reuse the main native tool pipeline

    A development checkpoint tag rather than a release, but it names the direction plainly: subagents call tools through the same native pipeline as the main agent instead of a parallel path. Nothing here is installable yet.

    View source ↗
  2. 28d ago

    v0.8.4: fix: persist chain-of-thought duration in thread metadata (#8474)

    Chain-of-thought duration is now computed at stream completion and written into thread metadata, so navigating away and back no longer resets it to a vague placeholder. A display bug fixed at the persistence layer rather than in the component.

    View source ↗
  3. 1mo ago

    v0.8.3: fix(csp): allow data:/blob: media so video uploads load (#8330)

    A missing media-src directive in the content security policy blocked data: and blob: media, so uploaded video would not load. Mirroring the existing img-src rule fixes it.

    View source ↗
  4. 2mo ago

    llama.cpp auto-fit disabled by default, context length set to 8192

    Auto-fit is off by default and context length defaults to 8192 — a change to what every local model does out of the box, trading automatic memory sizing for predictable behavior. In a local-inference app, defaults like this shape the experience more than features do.

    View source ↗