Filigree

Local-first issue tracker designed for AI coding agents — SQLite, MCP tools, no cloud, no accounts.

What Is Filigree?

Filigree is a lightweight, SQLite-backed issue tracker designed for AI coding agents (Claude Code, Codex, etc.) to use as first-class citizens. It exposes 114 MCP tools so agents interact natively, plus a full CLI for humans and background subagents.

Traditional issue trackers are human-first — agents scrape CLI output or parse API responses. Filigree flips this: agents get a pre-computed context.md at session start, claim work with optimistic locking, and resume sessions via event streams without re-reading history. For Claude Code, filigree install wires up session hooks and a workflow skill pack so agents get project context automatically.

Filigree is local-first. No cloud, no accounts. Each project gets a .filigree/ directory (like .git/) containing a SQLite database, configuration, and auto-generated context summary. Installations support two modes: ethereal (default, per-project) and server (persistent multi-project daemon). Filigree 2.0 also adds a named Loom HTTP generation at /api/loom/* for federation-aware integrations while keeping the classic HTTP surface supported for existing callers.

Security boundary: Filigree does not encrypt, sandbox, harden, or secure stored project data beyond ordinary filesystem permissions and standard HTTPS transport if you put it behind HTTPS yourself. Do not use Filigree for secure, regulated, confidential, or business-sensitive data.

Key Features

• MCP server with 114 tools — agents interact natively without parsing text
• Full CLI with --json output for background subagents and --actor for audit trails
• Loom HTTP generation — stable /api/loom/* contracts with classic compatibility for existing integrations
• Claude Code integration — session hooks inject project snapshots at startup; bundled skill pack teaches agents workflow patterns
• Workflow templates — 24 issue types across 9 packs with enforced state machines
• Dependency graph — blockers, ready-queue, critical path analysis
• Hierarchical planning — milestone/phase/step hierarchies with automatic unblocking
• Atomic claiming — optimistic locking prevents double-work in multi-agent scenarios
• Agent handoff tools — claim leases, stale-claim discovery, observation triage, shared file annotations, and scanner/finding workflows
• Bundled scanners — opt-in Codex and Claude scanner registrations with prompt packs for security, architecture, language, and quality reviews
• Pre-computed context — context.md regenerated on every mutation for instant agent orientation
• Web dashboard — real-time project overview with Kanban drag-and-drop, Graph v2 dependency exploration, Files/Health views, and optional multi-project server mode
• Minimal dependencies — just Python + SQLite + click (no framework overhead)
• Session resumption — get_changes --since <timestamp> to catch up after downtime

How It Works

Filigree gives one project three ways into a single SQLite database — agents through MCP tools, you through the CLI, and a browser through the dashboard. Every mutation regenerates context.md, which the session hook injects back into the agent at startup.

flowchart TD
    Agent["AI coding agent<br/>(Claude Code, Codex)"]
    Human["You + scripts"]
    Browser["Browser"]

    MCP["MCP server<br/>(114 namespaced tools, stdio)"]
    CLI["filigree CLI<br/>(--json, --actor)"]
    Dash["Web dashboard<br/>(localhost:8377)"]

    DB[("SQLite database<br/>.filigree/filigree.db")]
    Ctx["context.md<br/>(pre-computed orientation)"]
    Loom["Loom HTTP /api/loom/*<br/>(optional federation)"]

    Agent -->|"calls tools over stdio"| MCP
    Human -->|"runs commands"| CLI
    Browser -->|"views, drags Kanban"| Dash

    MCP -->|"reads & writes issues"| DB
    CLI -->|"reads & writes issues"| DB
    Dash -->|"reads & writes issues"| DB

    DB -->|"regenerates on every mutation"| Ctx
    Ctx -->|"injected at session start"| Agent

    DB -.->|"serves federation peers"| Loom

Loading

The work loop is the same whether you drive it from the CLI or an agent drives it through MCP tools: orient at session start, claim an issue atomically, do the work, close it — and the next orientation is already regenerated.

sequenceDiagram
    participant Dev as You / AI agent
    participant Fg as Filigree
    participant DB as .filigree/ (SQLite)

    Note over Dev,Fg: Session start
    Dev->>Fg: filigree session-context
    Fg->>DB: read ready / in-progress / critical path
    DB-->>Dev: project snapshot

    Dev->>Fg: filigree start-next-work --assignee me
    Fg->>DB: atomic claim + transition to working status
    DB-->>Dev: claimed issue (optimistic lock prevents double-work)

    Note over Dev: do the work, commit

    Dev->>Fg: filigree close <id>
    Fg->>DB: transition to terminal status
    DB->>DB: regenerate context.md
    DB-->>Dev: next ready work

Loading

Each issue type enforces its own state machine. Tasks are simple (open → in_progress → closed); features add review (proposed → approved → building → reviewing → done); bugs add a triage front and a hard verification gate, so a fix cannot close without a recorded fix_verification. The bug lifecycle below shows why "ready" is not the same as "startable" — a bug enters at triage, which has no single hop into work.

stateDiagram-v2
    [*] --> triage
    triage --> confirmed
    triage --> wont_fix
    triage --> not_a_bug
    confirmed --> fixing
    fixing --> verifying : requires fix_verification
    verifying --> closed : hard gate (fix_verification required)
    verifying --> fixing : kicked back
    confirmed --> wont_fix
    wont_fix --> [*]
    not_a_bug --> [*]
    closed --> [*]

Loading

Quick Start

pip install filigree        # or: uv add filigree
cd my-project
filigree init               # Create .filigree/ directory
filigree install             # Set up MCP, hooks, skills, CLAUDE.md, .gitignore
filigree create "Set up CI pipeline" --type=task --priority=1
filigree ready               # See what's ready to work on
filigree update <id> --status=in_progress
filigree close <id>

Installation

pip install filigree                     # CLI + MCP server + Web dashboard

Or from source:

git clone https://github.com/tachyon-beep/filigree.git
cd filigree && uv sync

Entry Points

Command	Purpose
filigree	CLI interface
filigree-mcp	MCP server (stdio transport)
filigree-dashboard	Web UI (port 8377)

Claude Code Setup

filigree install configures everything in one step. To install individual components:

filigree install --claude-code   # MCP server + CLAUDE.md instructions
filigree install --codex         # Codex MCP config via runtime autodiscovery
filigree install --hooks         # SessionStart hooks (project snapshot + dashboard auto-start)
filigree install --skills        # Workflow skill pack for agents
filigree doctor                  # Verify installation health

The session hook runs filigree session-context at startup, giving the agent a snapshot of in-progress work, ready tasks, and the critical path. The skill pack (filigree-workflow) teaches agents triage patterns, team coordination, and sprint planning step by step.

Dashboard Authentication Scope

The dashboard is local-first and assumes loopback/local filesystem trust by default. Setting FILIGREE_FEDERATION_API_TOKEN enables bearer-token authentication only for federation and agent-ingest surfaces; the older FILIGREE_API_TOKEN is still accepted as a backward-compatible fallback. The token value is never reported by /api/health, but the health payload does report which auth scope is enabled.

Route class	Authentication
Dashboard UI (/)	Open under the local loopback trust boundary
Classic dashboard API (/api/issues, /api/issue/{id}, /api/health)	Open under the local loopback trust boundary
Federation and scanner ingest (/api/loom/*, /api/scan-results, /api/observations, /api/v1/scan-results)	Bearer token when FILIGREE_FEDERATION_API_TOKEN or fallback FILIGREE_API_TOKEN is set
MCP HTTP endpoint (/mcp, /mcp/*)	Bearer token when FILIGREE_FEDERATION_API_TOKEN or fallback FILIGREE_API_TOKEN is set

Why Filigree?

Filigree is designed for a specific niche: local-first, agent-driven development. It is not a replacement for GitHub Issues or Jira.

Is Filigree suitable for my project? Use Filigree when your priority is plug-and-play agent coordination: install it, run filigree init and filigree install, open your agent, and immediately get project context, tickets, scanner findings, and local workflow tools. Do not use Filigree when the issue database, comments, scans, or integrations would contain secrets, regulated data, customer data, confidential business information, or anything that must be encrypted or access-controlled beyond normal local filesystem protections and HTTPS transport you provide yourself.

Feature	Filigree	GitHub Issues	Jira
Agent-native MCP tools	Yes	No	No
Works offline, no account needed	Yes	No	No
Enforced workflow state machines	Yes	Limited	Yes
Dependency graph + critical path	Yes	Limited	Yes
Structured queries and filtering	Yes	Yes	Yes
Multi-user cloud collaboration	No	Yes	Yes
Integration ecosystem (CI, Slack)	No	Yes	Yes

When NOT to Use Filigree

Filigree is designed for one niche well. It is the wrong tool if you need:

Team collaboration across machines. Filigree has no cloud sync, no accounts, and no network-accessible API beyond localhost. If your team of humans needs to file bugs, assign tickets, and comment across different machines, use GitHub Issues, Linear, or Jira.

Integration with your existing toolchain. Filigree does not connect to CI pipelines, Slack, PagerDuty, or third-party services. If your workflow requires automated ticket creation from alerts or Slack-based triage, Filigree will not fit without custom scripting.

A persistent project record outlasting the repository. Your .filigree/ directory lives with your project. If you need an audit trail that survives repository deletion or is accessible after the project ends, use a hosted service.

Multi-project portfolio management. The web dashboard supports switching between local projects, but Filigree has no cross-project reporting, resource allocation, or roadmap views. It tracks tasks, not portfolios.

Mobile or browser-based access. The dashboard runs on localhost. If stakeholders need to read or file issues from their phone or a machine where the project is not checked out, Filigree is not the right choice.

Secure or sensitive data. Nothing in Filigree is encrypted or secured beyond ordinary local filesystem protections and standard HTTPS if you provide it. Do not use this system for secure or sensitive data.

The sweet spot: one developer or agent team, one project, offline or airgapped, where you want structured workflow enforcement and agent-native tooling without standing up external infrastructure.

Documentation

Document	Description
Getting Started	5-minute tutorial: install, init, first issue
CLI Reference	All CLI commands with full parameter docs
MCP Server Reference	114 MCP tools for agent-native interaction
Federation Contracts	Classic and Loom HTTP generation contracts
Workflow Templates	State machines, packs, field schemas, enforcement
Agent Integration	Multi-agent patterns, claiming, session resumption
Python API Reference	FiligreeDB, Issue, TemplateRegistry for programmatic use
Architecture	Source layout, DB schema, design decisions
Examples	Runnable scripts: multi-agent, workflows, CLI scripting, planning

Priority Scale

Priority	Label	Meaning
P0	Critical	Drop everything
P1	High	Do next
P2	Medium	Default
P3	Low	When possible
P4	Backlog	Future consideration

Full definitions: Workflow Templates — Priority Scale

Development

Requires Python 3.11+. Developed on 3.13. Node.js 24 is also required for the Node-backed static dashboard pytest tests and dashboard JavaScript quality gates.

git clone https://github.com/tachyon-beep/filigree.git
cd filigree
uv sync --group dev

make ci              # ruff check + mypy strict + pytest with 85% coverage gate
make lint            # Ruff check + format check
make format          # Auto-format with ruff
make typecheck       # Mypy strict mode
make test            # Pytest
make test-cov        # Pytest with coverage (fail-under=85%)

Key Conventions

• Ruff for linting and formatting (line-length=140)
• Mypy in strict mode
• Pytest with pytest-asyncio for MCP server tests
• Coverage threshold at 85%
• Tests in tests/, source in src/filigree/

Acknowledgements

Filigree was inspired by Steve Yegge's beads project. Filigree builds on the core idea of git-friendly issue tracking, focusing on MCP-native workflows and local-first operation.

License

MIT — Copyright (c) 2026 John Morrissey