Agent Adapter

Reuse Python tools across agent frameworks with one execution pipeline for approvals, budgets, and traces.

Stars
★ 13
Last updated
3d ago
License
Apache-2.0
Primary language
Python

At a glance

How it runs
Library / SDKFrameworkCLI
Works with
Universal · cross-platformOpenAI API · Claude APICodex · Claude Code (Partial support)
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Python 3.11+PydanticShell / CLINetwork accessLocal filesystemMCP Server
Typical use
Engineering teams moving between LangGraph and CrewAI who want to reuse the same Python tools.
Not a fit if
  • Teams seeking a ready hosted service
  • Teams that need to write core tools directly in TypeScript
Website
agent-adapter.veloceai.in ↗ set on GitHub
Maintenance
Commits in the last 30 dayslast commit 2026-11-18latest release v1.2.0 (2026-11-18)
Source review
72/100 · Some gaps

What does this agent do, and when should you use it?

Agent Adapter is a framework-independent Python SDK that requires Python 3.11 or later. Developers register ordinary Python functions with @aa.tool; its executor validates arguments and applies configured approvals, budgets, caching, retries, and cost tracking. Optional adapters connect frameworks such as LangChain, CrewAI, and Google ADK, as well as OpenAI, Gemini, and Anthropic model APIs; tools can also be exposed over MCP, HTTP, A2A, or AG-UI. It includes Flow graph orchestration and the run_tool_loop agent loop, with events, ledger records, and OpenTelemetry traces as outputs. It can be embedded as a library in a service or script, or run as an HTTP or MCP server with a local run viewer.

Register Python functions with @aa.tool or registry.wrap(); schemas are checked at startup. Calls through aa.call, aa.acall, or Executor.run validate arguments before running configured guards, budget checks, approvals, caching, retries, and timeouts. run_tool_loop drives a model port through repeated tool calls until it returns an answer; Flow and FlowRunner execute workflows with routing, parallel branches, merges, and human pauses. Framework adapters, model APIs, MCP, and HTTP interfaces route calls through the same executor pipeline. Outputs include tool results, call costs, events, and traces; the optional viewer displays run records, flow graphs, timelines, and a chat interface.

  1. Engineering teams moving between LangGraph and CrewAI who want to reuse the same Python tools.
  2. Service teams that need human approval for high-risk actions such as refunds and want to persist pending calls in SQLite for later resumption.
  3. Application developers who need per-run call limits or spending caps with recorded tool costs.
  4. Teams exposing Python tools to clients in other languages through MCP or HTTP.
  5. Developers building routed or parallel agent workflows with human pauses and a local run viewer for inspection.

How do you install or deploy this agent?

Requires Python 3.11 or later. The base package depends on Pydantic; install the relevant optional extra for model ports or framework adapters. The README does not mandate a model provider or give a universal credential requirement; credentials depend on the provider selected.

How do you use this agent?

Register a tool in the default registry and call it through the default executor. If you use a model or provider SDK that requires credentials, configure those according to that provider; the basic invocation below requires no model API key.

What are this agent's strengths and limitations?

Pros
  • The same Python tools can be reused through multiple framework adapters, while approvals, budgets, and execution rules stay in one executor.
  • DeferredApproval can persist pending calls in SQLite for later resumption; approval tokens are tied to a specific call and used once.
  • Includes Flow, run_tool_loop, MCP and HTTP interfaces, and a local run viewer, spanning tool execution through workflow inspection.
  • The core depends only on Pydantic; model and framework integrations are optional extras imported lazily.
Limitations
  • Requires Python 3.11+; core tools are Python functions, while other languages connect through MCP or HTTP clients.
  • Frameworks, model providers, persistence stores, and guard backends generally need separate extras, and some integrations require their own environment.
  • Online model use requires provider credentials and may incur provider charges; the README gives no unified cost estimate.
  • Configuration and deployment depend on the selected features; the README lists several backends and stores that still need to be chosen and configured.

How does this agent compare with similar options?

Compared with using the native tool formats of LangChain, CrewAI, or ADK directly, Agent Adapter keeps tool registration and execution outside the framework and provides adapters for those frameworks. Applications that do not need an agent framework can use its executor, run_tool_loop, or Flow directly.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Agent Adapter This agent 72 · Some gaps Library / SDKFree + model costs ★ 13 3d ago Python OpenAI API · Claude API
AG2: Open-Source AgentOS 75 · Good Library / SDKFree + model costs ★ 5.1k 1d ago Python OpenAI API
mcp-agent 60 · Some gaps FrameworkFree + model costs ★ 8.6k 9mo ago Python Claude.ai · OpenAI API · Claude API
Pocket Flow 45 · Major gaps Library / SDKFree + model costs ★ 11k 3mo ago Python —

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Some gaps
72/ 100 5-point scale 3.6 / 5
This review covers the source as of Nov 13, 2026. The project has shipped 1 stable release(s) since, which have not been re-reviewed. The score and verdict describe the reviewed version only. Recent releases ↓
Trust 19/29
Reliability 9/14
Adaptability 15/18
Convention 15/18
Effectiveness 9/13
Verifiability 5/8
Why each dimension lost points
Trust19 / 29 · 3.3/5

Security docs and code show fail-closed defaults (confirm calls denied without an approval port), one-use approvals bound to run/call/argument digest, HTTP server requiring auth by default, log redaction and hidden crash text; approval and sandbox design are fairly complete. least_privilege scores 2: sandbox docs explicitly state the host kernel is not isolated, LocalSubprocessSandbox is dev-only, and code_tool depends on approval-port strength. data_flow_transparency scores 2: event catalogue and OTel exist, but no full data-flow map or per-external-call data destination list. sensitive_data_handling scores 2: PIIGuard and redaction exist, but Presidio/NeMo are optional and need extra models, default coverage unproven. dependency_security scores 2: dependencies have bounds and an h11 CVE is mentioned, but no SBOM, lockfile or vulnerability-scan evidence. external_effects scores 2: budgets, rate limits, timeouts and side-effect confirmation exist, but retry/timeout side effects (sync tool keeps running in its thread) are only documented as a caution. rollback scores 1: only approval resume and audit log; no tool side-effect rollback or compensation. source_attribution scores 2: LICENSE/NOTICE and author info exist, but publisher identity is unverified and attribution is only partly checkable.

Reliability9 / 14 · 3.2/5

self_consistency scores 2: README, AGENTS.md, SKILL.md and pyproject broadly agree on version, Python requirement and install, but README says 21 frameworks while AGENTS.md lists more, and some examples depend on files not shown. dependency_availability scores 2: optional dependencies have explicit ranges and own-environment conflicts are documented, but no lockfile or resolved tree is provided, so installability cannot be statically confirmed. failure_messages scores 2: error hierarchy (AdapterError, ToolUsageError, ConfigError, etc.) and CLI exit codes are defined, but some error texts and recovery advice appear only in docs, not all shown in code samples.

Adaptability15 / 18 · 4.2/5

audience_and_scenarios scores 3: README, SKILL.md and AGENTS.md cover scripts, services, frameworks, multi-language and no-framework scenarios with clear audiences. capability_boundaries scores 3: security docs explicitly list what the adapter does not provide (sandboxing, secrets store, content moderation, per-user rate limiting), boundaries are clear. trigger_precision scores 2: SKILL.md description triggers broadly (any agent tool need) and may over-trigger; no exclusion conditions given. environment_fit scores 2: supports Python 3.11–3.13 and many backends, but some frameworks need separate environments and Windows local sandbox has known limits; cross-platform fit not fully verified.

Convention15 / 18 · 4.2/5

information_architecture scores 3: README is layered, docs organized by get-started/how-to/concepts/reference, project layout explained. install_notes scores 3: install commands, extras table, own-environment notes and doctor check are complete. naming_stability scores 2: an API stability policy link exists, but no concrete versioned naming rules or deprecation process shown in the sample. examples_and_faq scores 3: many runnable examples, code tour and demo links. known_limitations scores 3: security and sandbox docs clearly list non-provided capabilities and residual risks. license scores 3: Apache-2.0 LICENSE and NOTICE present, pyproject uses SPDX expression. versioning_changelog scores 2: version number and CHANGELOG link exist, but the sample does not include CHANGELOG content or semantic-versioning commitment details. maintenance_responsibility scores 1: publisher is unverified Veloce AI, no maintainer list, security contact or SLA, only an author field.

Effectiveness9 / 13 · 3.5/5

output_usability scores 2: tool outputs are structured (result, cost, trace), but users must wire the executor and approval port themselves; out-of-the-box usability is moderate. marginal_value scores 2: a unified executor and cross-framework adapter have clear value, but competition is strong and differentiation depends on unverified integration quality. cost_benefit scores 2: core depends only on pydantic and extras are opt-in, but many framework integrations add maintenance and conflict cost, and no performance or cost benchmarks are provided.

Verifiability5 / 8 · 3.1/5

claim_traceability scores 2: most claims trace to README, security docs or code, but numbers like 90% coverage and 21 frameworks lack reproducible verification commands or reports in the sample. cross_source_corroboration scores 2: README, AGENTS.md, SKILL.md and pyproject corroborate version and install, but tests and CI only show workflow files, not actual test results. fact_inference_separation scores 2: docs separate 'adapter guarantees' from 'your app must', but some security claims (e.g. fail closed) rely on code paths not shown, so static review cannot fully separate fact from inference.

Risks and how to mitigate them
  • Publisher identity is unverified; maintenance responsibility and security contact are unclear, so verify independently before production use.
  • Sandbox does not isolate the host kernel and LocalSubprocessSandbox is dev-only; code_tool safety depends entirely on approval-port strength.
  • A timed-out sync tool may keep running in its thread and a retry can repeat side effects; documented but no automatic compensation.
  • Many optional dependencies, some needing separate environments; no lockfile or SBOM, so install conflicts and vulnerability status must be verified independently.
  • Claims such as 90% coverage and 21 frameworks lack reproducible verification reports in the sample and cannot be confirmed by static review.
Evidence confidence: Low Reviewed Nov 13, 2026 Reviewed revision 8152babd9977 New commits since this review; the score may not cover them
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/agent-adapter.svg)](https://followagents.com/en/agents/agent-adapter)
<a href="https://followagents.com/en/agents/agent-adapter"><img src="https://followagents.com/badge/agent-adapter.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v1.2.0 v1.2.0: VoxCat voice, shared board, template documents
    Release notes excerpt
    Voice, video and robots, agents that share what they find, and documents filled from many model calls. ```bash pip install -U agent-adapter pip install -U "agent-adapter[voice-cloud]" # hosted speech-to-text and text-to-speech ``` Highlights - VoxCat, our own voice pipeline (agent_adapter.voice). Phone calls (Twilio, Telnyx, Plivo, Vonage), browser calls over WebRTC, and the local mic. Spoken approvals before side effects, and nothing is done twice when the caller interrupts. It is fast with any speech-to-text and text-to-speech, open models included: its own end of turn (`Eager …
    View on GitHub ↗
  2. v1.1.1 v1.1.1: DSPy 3.5 support
    Release notes excerpt
    DSPy 3.5 support. Agent Adapter now works with DSPy 3.5 and 3.4. ```bash pip install -U "agent-adapter[dspy]" ``` Added - dspy_compress_prompt(prompt, policy) in agent_adapter.integrations.dspy, and the "dspy" compression format for the lm15 messages in a DSPy 3.5 Prompt. It takes what Adapter.format returns on DSPy 3.5 or 3.4 and gives back the same kind. Fixed …
    View on GitHub ↗
  3. v1.1.0
    Release notes excerpt
    The first public release: open source under Apache-2.0 and published on PyPI. Added - Guardrails. PIIGuard: a dependency-free guard hook for emails, phone numbers, cards (Luhn), IBANs (mod-97) and API keys or secrets. It blocks, redacts tool results and model replies, or sends flagged tool calls to one-use approval (confirm=guard.confirm). It fails closed and never logs content. - CallableGuard(fn) for any guard service; PresidioGuard (extra presidio) and NemoGuard (extra nemo); ContentGuard, TextGuard and GuardVerdict to write your own. …
    View on GitHub ↗

FAQ

Do I have to choose an agent framework?
No. You can call Executor directly, use run_tool_loop, or run workflows with FlowRunner; framework integrations are optional.
Is the software free, and do model calls cost money?
The project is licensed under Apache-2.0. If you choose an online model provider, its fees and credentials depend on that provider; the README gives no universal pricing.
How can a risky tool call wait for human approval?
Mark the tool confirm=True and configure an approval port such as DeferredApproval.sqlite. A pending call can be resumed later, and an approval cannot be reused.
Can clients in other languages call these Python tools?
The README documents MCP and HTTP servers plus a TypeScript client, and lists client examples for Python, C#, Go, Rust, Ruby, Java, Kotlin, Swift, and PHP.
Is Agent Adapter open source?
Yes. The source code is on GitHub (VeloceAI/agent-adapter) under the Apache-2.0 licence.
Is Agent Adapter safe to use?
FollowAgents' FARS review (Nov 13, 2026) scored Agent Adapter 19/29 on trust and safety. None of the eight safety controls the review checks for was missing outright. The repository has changed since that review.
Is Agent Adapter still maintained?
Yes. The latest commit on GitHub is from 2026-11-18, within the last 30 days. The latest release is v1.2.0 (2026-11-18).
Who makes Agent Adapter?
Agent Adapter is published on GitHub by the account VeloceAI. FollowAgents has not verified which person or company is behind that account.
What is Agent Adapter's official website?
The project's GitHub repository lists https://agent-adapter.veloceai.in as its website. That link is set by the repository owner; the source code itself is on GitHub (VeloceAI/agent-adapter).
Was this page helpful?
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related concepts