Skip to content

Latest commit

 

History

1,058 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Deepr

CI License: Apache 2.0 Python 3.12+ Version

Persistent domain experts built from bounded, auditable research.

Deepr turns research into durable beliefs, gaps, contradictions, confidence, citations, provenance, and outcomes. It prefers owned local models, then proven subscription quota, with no automatic fallback to a paid API.

Build a durable expert fleet

Deepr Expert Hub showing the 25-expert flagship roster with portraits, standpoints, positions, studied findings, and retained sources

The pictured maintainer fleet keeps its selected 25 flagship experts in focus while preserving its complete roster. Flagship membership is user-curated local state, not a 25-expert clean-install guarantee. Each card exposes the expert's standpoint, positions, studied findings, retained sources, and readiness from durable state.

Turn evidence into reusable judgment

A Deepr temporal knowledge graphs expert profile showing inspectable claims, confidence, source counts, and knowledge domains

Expert profiles keep claims, confidence, source lineage, gaps, decisions, and history inspectable after a research run ends. The structured belief store is authoritative; reports, digests, and portable exports are regenerable views.

Keep the spend boundary visible

A $200 cumulative Deepr wallet with a separate $4 job ceiling, exact drawdown, no automatic refill, and provider hard-stop status shown separately

Local and eligible plan-quota work stay at $0 marginal API cost. Metered work never becomes an automatic fallback and cannot turn the local wallet into an open check.

Capacity

Class Current posture
Local Ollama Preferred for expert setup, maintenance, evaluation, and consultation after endpoint ownership is proven. Records $0 and does not consume wallet capacity.
Plan quota Uses an existing subscription only when authentication, tool confinement, remaining quota, and disabled paid overage are proven. Claude Code is the current executable adapter. Successful work records $0 at the margin.
Metered API No automatic fallback. The attended absorb path requires verified provider prepaid-no-overage or a hard provider ceiling, plus a cumulative Deepr wallet, a separate finite job ceiling, explicit confirmation, and a durable reservation. Other metered surfaces remain gated.

A wallet is one cumulative drawdown across provider calls and time windows, not a per-request allowance. The operator chooses its size, whether $2, $50, or $200, and every settled API dollar plus every active paid hold consumes it. There is no overdraft, automatic refill, or universal $2 job maximum. Every job still needs its own narrower confirmed ceiling.

Wallet funding is local Deepr authorization. It does not buy or verify provider credits. Provider-side prepaid credits or a provider-enforced hard stop with paid overage disabled are also mandatory for dispatch. Deepr verifies both layers and applies the tighter boundary. A soft budget alert on an open postpaid account is not a hard stop and remains execution-blocked. MCP, schedules, loops, consultation, and automatic routing cannot use wallet authority. No production provider account-control verifier ships in v2.50, so actual paid dispatch remains blocked even after local wallet funding. The wallet and attended transaction are ready for a future authenticated verifier; they are not a bypass for today's quarantine.

# Keep the active example's independent calendar ceiling conservative.
# In .env: DEEPR_MAX_COST_PER_MONTH=5.00
deepr budget credits add --amount 200.00 --reason "Bound this API campaign"
# This remains blocked unless capacity status proves a provider hard stop.
deepr expert absorb "My Domain Expert" --file report.md --api --budget 4.00
deepr budget status
deepr costs show
deepr costs doctor
deepr budget credits clear

See Capacity and Cost for billing reconciliation and the complete no-surprise-spend contract.

Install

Windows PowerShell:

powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/blisspixel/deepr/main/scripts/install.ps1 | iex"

Linux and macOS:

curl -fsSL https://raw.githubusercontent.com/blisspixel/deepr/main/scripts/install.sh | bash

Installers use the latest verified GitHub Release wheel in an isolated pipx environment. PyPI publication is not enabled.

Quick start

deepr init
deepr doctor --skip-connectivity
deepr capacity
deepr expert make "My Domain Expert" --local -d "The decisions this expert supports"
deepr expert consult "What should we decide next?" --expert "My Domain Expert"

See Quick Start, Supported Surface, and the MCP Agent Guide for workflows across the current 36 MCP tools. The MCP server implements the final 2026-07-28 protocol revision (stateless per-request negotiation, server/discover, subscriptions/listen, Streamable HTTP header and Origin validation) while still serving legacy initialize-era clients on both transports. Operators can prove the offline host-interop posture with deepr mcp conformance ($0, no network, no model).

The portability roadmap follows the published Agent Plugins 1.0.0 specification and containment contract. OKF export remains a derived view over Deepr's canonical belief and provenance stores. Export and $0 validation now target the hard bundle rules in the current Open Knowledge Format 0.2 specification, while import remains untrusted and verification-gated. Agent Plugin packaging is now available as a contained, stdio-only, capability-read-only package foundation under packages/deepr-agent-plugin. Its schemas and Agent Skill are validated offline, its spend ceilings are fixed at zero, and its archive build is byte reproducible. deepr mcp host-profile openclaw now emits a deterministic, config-only reference for stable OpenClaw v2026.7.1-2. It uses the exact ten-tool read-only catalog, explicit contained data roots, and zero primary and legacy spend ceilings. The command does not install OpenClaw, change host configuration, inspect credentials, or claim live validation. Independent host conformance, additional profiles, remote routes, steering, and external computer control remain gated in Roadmap.

Documentation

Development

uv pip install -e ".[dev,full]"
pytest tests/unit/ --ignore=tests/data -q
ruff check src/deepr/
ruff format --check src/deepr/
python scripts/check_file_sizes.py
python scripts/check_ratchets.py
python scripts/check_paid_api_boundaries.py

Do not run bare pytest: integration tests can contact real providers. The blocking unit suite requires at least 80 percent branch coverage.

License

Apache 2.0

About

Open-source, multi-provider research automation with persistent experts, eval-based routing, and cost guardrails.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

10 stars

Watchers

1 watching

Forks

Releases

Used by

Contributors

Languages