kbforge

Agent-first knowledge bases, forged from your systems of record.

The Open Knowledge Format (OKF) v0.2 standardizes the artifact at rest — markdown concept files, frontmatter, index.md, log.md. It says nothing about how those bundles get produced: how you pull from a wiki or a CMDB, how you tell a real change from an export timestamp jittering, how a claim stays traceable to its source, and how an update reaches main without a human losing an afternoon to review.

kbforge is the missing half — the production protocol.

"Agent-first" is a checkable claim, not a downstream hope. kbforge stays a producer — the agent connects over MCP, which kbforge doesn't own — but every publish is gated on four agent-facing artifact laws (facet well-formedness, link resolvability, anchor presence, freshness legibility), plus a projection↔files coherence check so nothing ships unvalidated. That gate is what puts the frontmatter, links, and provenance an agent's serving layer needs into the artifact. What each law enforces at full versus reduced strength (and the paths to full strength) is spelled out honestly in architecture.md §4.4.

Status

Beta. The full pipeline runs end to end, and every seam that can only be checked against a real service has a live suite (--run-live) rather than a mock.

• Sources — local_files and git_commits built in, no credentials. Any MCP server with a select tool and a read-by-id tool becomes a source through configuration alone via kbforge-mcp, live-tested against AWS Documentation and GitHub. Any database SQLAlchemy can reach becomes a source through configuration alone via kbforge-sql, one scoped query per source. Third-party connectors are discovered through an entry point with no change to kbforge.
• Core — canonicalization under a stability law, a replay-safe mirror and diff, change detection, the no-op rule, incremental sync via a real cursor, and cross-source grounding: one owning document per concept, cited alongside grounding documents from other systems.
• Gate — the §4.4 emit-side laws plus a projection↔files coherence check, so nothing ships unvalidated.
• Synthesis — a deterministic stub (default, no LLM) or an opt-in grounded LLM synthesizer (--synthesizer llm, via the kbforge[llm] extra).
• Publishers — dry-run built in; GitHub and GitLab opt-in via --publisher with the token from an env var. kbforge opens review requests and never merges.
• Reading a bundle back — kbforge-okfquery loads a published bundle into DuckDB.

Still 0.x, and the API can still move: docs/design/ holds specs for work that is designed but unbuilt, and sections of docs/architecture.md headed not built are specification rather than shipped code. The largest known gap is a first-party connector for a specific system of record — the capability exists through kbforge-mcp, but nothing ships preconfigured for Confluence, ServiceNow, or their kin.

Companion distributions

This repo is a uv workspace and ships four distributions, versioned independently. No companion is required to use kbforge, and none is installed with it.

They sit on opposite sides of the pipeline. kbforge-mcp is an ingest-side plugin that registers a connector entry point; okfquery registers nothing and imports no kbforge code at runtime, so it reads any OKF v0.2 bundle, kbforge-produced or not.

Quickstart

pip install kbforge
kbforge list                       # show available connectors

kbforge run \
  --connector local_files \
  --set path=./docs \
  --mirror .kbforge/mirror --out .kbforge/out --state .kbforge/state

Re-running with no source change is a no-op — no merge request is opened. Point --connector git_commits --set repo=. at a git repository to sync commit history incrementally instead. Config values are YAML-typed, so --set max_commits=50 is an integer and --set 'ignore_globs=[drafts]' is a list.

To synthesize real prose instead of the deterministic stub, install the LLM extra and select the synthesizer (config values are YAML-typed; the API key comes from an env var, never the CLI):

pip install "kbforge[llm]"
export OPENROUTER_API_KEY=...        # or point --llm-set api_base=... at a gateway
kbforge run --connector local_files --set path=./docs \
  --synthesizer llm --llm-set model=deepseek/deepseek-v4-flash \
  --mirror .kbforge/mirror --out .kbforge/out --state .kbforge/state

The synthesizer reaches models through a LiteLLM provider, so OpenRouter and a self-hosted LiteLLM gateway share one config path.

Publishing to GitHub or GitLab

The default publisher writes the proposal to a local directory. To open a real pull request or merge request instead, select a forge publisher and give it a repo. The token comes from an env var, never the CLI:

export GITHUB_TOKEN=...            # or GITLAB_TOKEN
kbforge run --connector local_files --set path=./docs \
  --publisher github --publish-set repo=acme/knowledge-base \
  --mirror .kbforge/mirror --out .kbforge/out --state .kbforge/state

Both publishers accept the same config: repo (required), base (default: the repo's default branch), base_path (a subdirectory, default: repo root), branch (default: sync/<system>), title, api_base (point it at GitHub Enterprise or a self-managed GitLab), and token_env.

kbforge maintains one long-lived sync branch and one open review request per source system: a later run force-updates that branch and edits the existing PR/MR rather than opening a second one. Three consequences worth knowing:

• Concepts deleted from the source are deleted from the target repo, provided the connector emits an explicit tombstone. Absence never implies a deletion.
• Manual commits on the sync branch are preserved while its review request is open — a later run builds on the branch rather than resetting it. Once no request is open, the next run rebuilds the branch from the default branch and those commits are gone. A hand edit to a concept kbforge later regenerates is overwritten by that regeneration either way.
• Close a kbforge review request only by merging it. The mirror advances on every successful publish, so the concepts a request carries are never re-proposed. Closing one unmerged discards its contents permanently: the target repo simply never gets them, and a published-then-abandoned deletion leaves the doc gone from the mirror, so no later run even sees it as a removal. To undo an abandoned request, reset both the mirror and the connector's cursor: delete the mirror directory and, in the state directory (--state), the connector's cursor-<connector-name>.json. Deleting the mirror alone does not work for an incremental connector — its cursor still points past the abandoned content, so the next kbforge_fetch returns only what changed since then, which can be little or nothing, and no re-proposal happens at all. Only once both are gone does a re-run re-propose everything from scratch.

kbforge never merges. No publisher has a merge method.

Design stance

The core ships zero credentialed connectors and zero CI logic. The two built-in connectors need no credentials and serve as references; real systems of record are plugins, discovered through the kbforge.connectors (and kbforge.publishers) entry-point group without editing kbforge — deployments are separate, private repositories. The interface is the product.

# in a third-party package's pyproject.toml — discovered automatically once installed
[project.entry-points."kbforge.connectors"]
myservice = "my_package:connector"

A complete worked example — a credentialed GitHub Issues connector (~160 lines) with token auth, pagination, and a real incremental cursor — is in examples/github-issues-connector/.

The pipeline order — fetch → normalize → mirror → diff → scope → synthesize → validate → publish — is deliberately not pluggable, and neither are the no-op rule or the never-auto-merge rule. Those are the trust guarantees; making them pluggable would make them optional. Plugins extend stages. They cannot reorder or remove them.

Documentation

• docs/architecture.md — package architecture, the Pluggy hookspecs, the connector protocol and its canonicalization laws, the fixed pipeline, the §4.4 emit-side laws, and the build sequence. Sections headed not built are specification rather than shipped code.
• docs/context/knowledge-base-design.md — the system kbforge was extracted from: an OKF knowledge base for application managers served over MCP, including the security model and a literature review.
• docs/design/2026-07-19-agentic-ingest-design.md — the roadmap for agentic fetch, the refresh model, and KB bootstrap.
• docs/design/2026-07-18-datacontract-bridge-design.md — how kbforge bridges to agentic-data-contracts via the OKF bundle (future, cross-project).
• docs/design/2026-08-08-okf-02-deferred-decisions.md — the OKF v0.2 families kbforge does not emit yet (verified, status: deprecated, stale_after, footnote attribution) and why each is a decision, not a backlog item.
• docs/design/2026-08-23-okfquery-design.md — why okfquery is a separate distribution rather than a kbforge query subcommand, its schema, and the two lossy mirror joins a reader has to know about.
• CHANGELOG.md — release history.

Related projects

kbforge is one of three contracts for agents, split by seam:

• ai-agent-contracts — the formal spine: resource, temporal, and lifecycle contracts (the seven-tuple kbforge maps onto).
• agentic-data-contracts — the consumption half for structured data: domain-driven governance enforced at query time. kbforge is the production half for unstructured knowledge; both independently converged on making freshness legible to the agent.

Not a context database

OpenViking and its kin sit in the serving row of the table above, not the production row. They ingest documents and expose them to an agent — OpenViking summarizes each one into retrieval tiers on write and serves them over a filesystem API and MCP. kbforge produces the documents such a system serves: an OKF bundle in a git repo is a valid input to one, so these compose rather than compete.

The difference shows on the second pull from a source that mostly did not change. A context database refreshes a watched resource by re-ingesting it wholesale — no diff, no changed-set, no proposal a human ever sees. kbforge canonicalizes first, so an export whose timestamps and ordering jitter reduces to no change at all: no LLM spend, no review request, no merge. Byte-level deduplication further downstream cannot substitute, because a rendered system-of-record export is rarely byte-identical across pulls even when nothing about it has meaningfully changed.

That one mechanism is why both the token bill and the review queue stay bounded on a corpus where most documents are stable — and it is what makes the human gate affordable rather than ceremonial. Reach for a context database when an agent needs to retrieve from a corpus; reach for kbforge when a corpus has to stay honest to a system of record that keeps changing, and someone has to be accountable for what it says.

Development

uv sync --all-extras --dev   # create the venv and install
prek install                 # ruff + ty on every commit
uv run pytest

The default suite never touches the network. Tests that call a real external service are marked live and skipped unless you pass --run-live.

The forge publishers have a live suite because their offline tests can only assert what we meant to send — a real forge is the only thing that can say the intent was right. It needs a throwaway repo on each forge and the two CLIs (gh, glab) authenticated; each run writes under a fresh live/<run-id>/ prefix, so nothing accumulates and no repo is ever deleted.

That suite covers two levels. Most tests hand-build a ProposedChange and pin the publisher adapters. One drives kbforge.pipeline.run end to end against a real forge and pins the composition — that an unchanged source is still a no-op, and that a later run accumulates into the same review request instead of rebuilding the branch and dropping concepts an earlier run put there. Those three pieces of state (mirror, cursor, open-request detection) can each be correct while the whole is not, which is exactly how the 0.3.0 data-loss bug passed a clean offline suite.

GITHUB_TOKEN=$(gh auth token) \
GITLAB_TOKEN=$(glab config get token --host gitlab.com) \
KBFORGE_LIVE_GITHUB_REPO=you/kbforge-live-test \
KBFORGE_LIVE_GITLAB_REPO=you/kbforge-live-test \
uv run pytest tests/test_forge_live.py --run-live

License

MIT