What are context sources?
Context sources are reference materials you attach to a session. The AI facilitator reads them during each participant conversation — so instead of working only from the session goal and prior discussions, it can draw on documents you’ve provided, summaries from related sessions, or live data from an MCP server. Up to 20 sources can be attached per session. Adding sources is a Pro feature. Sources can be attached to a single session or to a whole project, where they apply to every session it contains. Files, sessions and MCP servers work at either level; GitHub repositories are project-level only. See Managing projects for the project-level view.Source types
Files
Upload a document — a brief, a research report, a product spec, or any plain text or Markdown file. Short files (under 50 KB) are read in full on every turn. Longer files are chunked and indexed; the facilitator retrieves the most relevant passages based on what’s happening in the conversation. Use files when you want the facilitator to stay anchored in specific written material — for example, grounding a feedback session in an existing proposal, or keeping a planning session connected to a strategy document.Meeting transcripts
A meeting transcript is still a file, but Harmonica reads it differently: it recognises the speaker turns and tracks who said what, instead of treating the whole thing as one continuous document. Exports from Zoom or Meet work as they are, and a long transcript is chunked like any other large file. Detection happens automatically on upload. If you want to be explicit, or Harmonica reads it as a plain document, use Mark as transcript on the source in the sources list to set it yourself. Transcript sources carry a (meeting) label wherever they appear, so you can see at a glance which of your material is a conversation and which is a document. Marking a file as a transcript also makes it usable ground for a project’s opinion map and its seed statements, so a recorded discussion can feed a public page the same way a written source can.Marking a source as a transcript after it has been uploaded updates its label, but it does not re-read the file. Anything already extracted keeps the reading it was given when the file was first ingested.
Sessions
Reference another Harmonica session by ID. The referenced session’s summary is attached to this session as context. This is useful for cross-pollination: a second session can build on what the first group said, even if the participants are completely different. The snapshot is taken at the time you add the source. If the referenced session’s summary changes later, use the Refresh action on the source to pull in the updated version.GitHub repositories
A Team OS project can connect a GitHub repository, and Harmonica reads that repository’s markdown so every session in the project can draw on it. This is a project-level source, not a session-level one: connect it once and it applies across the project. Connect it from the project’s Memory record card, using Connect a repository. You install Harmonica’s GitHub app on the repositories you want to expose, then pick one from the list — Harmonica never asks you to paste a repository URL or a token. What it reads. All markdown files in the repository by default. You can narrow that to specific folders, picked from a list of the repository’s actual folders rather than typed as a path. Files are chunked and indexed like any large file upload, so the facilitator retrieves the passages relevant to the conversation rather than reading the whole repository each turn. What it tells you. The source shows how many files and chunks it actually read, and how many it skipped with the reason — over the size budget, or individually too large. A repository that is bigger than the budget is read partially and says so, rather than quietly dropping content. Keeping it current. Use Re-read on the source to pull the repository again after it has changed. If a read fails, the source says so instead of silently serving stale content.Disconnecting a repository severs the connection and its indexed content. It does not touch the entries already in the project’s memory record — anything the record has accumulated stays.
MCP servers
Attach an MCP (Model Context Protocol) server via its URL. The facilitator connects to the server live on each turn and can call its tools during the conversation. Use this for real-time integrations — a knowledge base, a live data feed, or a custom tool that should be available to the facilitator mid-session.MCP sources require a publicly reachable server URL. The facilitator connects on each participant turn, so latency and availability of your MCP server affects conversation speed. After five consecutive failures, a source is marked unhealthy and skipped until you reconnect it.
Attach a context source
1
Open Session Design
From your session results page, open the session overview panel and navigate to the Sources tab.
2
Click Add source
Click Add source. A dialog appears with three tabs: File, Session, and MCP.
3
Choose a source type and fill in the details
- File
- Session
- MCP
Select a plain text or Markdown file from your device. Give it a title so you can identify it in the sources list. Files over 50 KB are chunked automatically — you don’t need to do anything differently.
4
Confirm
Click Add. The source appears in the sources list immediately. The facilitator uses it in new conversations from this point forward.
Manage existing sources
From the Sources tab, each attached source shows its title, type, and status.- Remove — detaches the source. Existing conversations already in progress are not affected.
- Refresh (session sources only) — updates the attached summary from the referenced session. Useful if that session has produced new results since you originally attached it.
- Mark as transcript (file sources only) — tells Harmonica to read the file as a conversation between speakers rather than as a document. See Meeting transcripts.
How the facilitator uses context sources
Context sources are read at the start of each conversation turn, before the facilitator generates its response. The order of priority (from highest to lowest) in the facilitator’s view is:- The session’s core fields (goal, context, critical information)
- Any organizational context you’ve set at the account level
- Session-level context sources (files, sessions, MCP)
- Scratchpad themes accumulated from prior conversations
Related guides
Session settings
Other per-session configuration options
Chains
Using context from one step to inform the next
Session results
Reviewing what the facilitator surfaced
Templates
Starting from a pre-built session design